• Lumo 2.0 release rebuilds Proton's privacy-first assistant with reasoning modes, image generation/recognition, cited live web search, and persistent memory
• The privacy stack mixes cryptography and policy: zero-access encryption protects stored chats and images, while inference-time protection relies on Proton's no-logs/no-training promises that have held in the past
• Proton's Lumo 2.0 Lite and Lumo 2.0 Max score 127% and 240% higher than Lumo 1.4 on the Artificial Analysis Intelligence Index, making them close in on last-generation frontier AI models

Proton has revealed Lumo 2.0, its updated AI alternative to ChatGPT, Claude, and Gemini, focusing on privacy first and foremost, a distinctly different approach from most of its competition.

The new update is not only smarter than its predecessor at what it does, but also brings a host of new capabilities: reasoning modes, image generation and recognition, live web search with citations, persistent memory, and customizable assistants.

Lumo 2.0 looks to do all of this while leveraging zero-access encryption, no-logs, no-training, a pitch that makes it appealing to privacy-focused consumers, many of whom are already customers for its VPN product lineup.

Upgrades, multiple models and faster performance

The biggest upgrade to Lumo 2.0 is that it is now multimodal, allowing it to glean information and cross-check a variety of sources without often forcing the user to defer to other AI engines for most tasks.

Proton cites a 76% faster speed for 'everyday queries' while conceding that complex tasks still take a considerable amount of time.

Users can also leverage "Custom Lumos" or purpose-built assistants that retain instructions in memory while still maintaining the encryption promise that Lumo offers, allowing users to avoid starting from scratch each time they have a query to address.

Users can use either the fast, general-purpose Lite model for everyday queries and defer to the more complex Max model for demanding work, or use Fast and Thinking modes, which offer twice the context window of its predecessor for larger workloads and greater coherence with more complex asks.

Pricing spans a free tier for what Proton calls everyday private use, a $12.99-per-month Lumo Plus plan with unlimited chats, Projects, advanced image generation and access to the most capable models, and a $14.99-per-user Lumo Professional tier for teams.

Lumo is also available to business users and offers the same upgrades discussed above, making it a significantly more powerful and smarter AI tool than it was when we last reviewed it at TechRadar.

It is important to note that while Lumo 2.0 is a huge upgrade versus its older 1.4 version, it does not come as close to frontier models as Proton might want to make it appear: its model scores a 51 on the Artificial Analysis Intelligence Index which sees current frontier models clock in as high as 59 (GPT 5.6 Sol Max) or 60 (Claude Fable 5) versus its own comparisions that show it much closer to older frontier models such as GPT 5.5 and Claude Opus 4.8.

This is not entirely surprising, given that one can find the underlying tech Lumo uses in its privacy policy. Proton states that it uses a mix of Qwen 3.5, GLM 5.2, Image-Turbo, and FireRed-Image-Edit-1.1, with GLM 5.2's scores roughly identical to the numbers it cites currently.

Despite its limitations versus newer frontier AI models, Lumo 2.0 arguably remains the most privacy-focused approach to AI available to most end users currently, and it comes considerably closer than its predecessor at what is an increasingly uphill task of late: offering a competitive privacy-centric alternative to billion-dollar proprietary AI models built by the likes of Google, Anthropic, and OpenAI.

It's hard to get away from AI these days, with the technology finding its way into our apps, devices, and entertainment — but according to a survey carried out by TechRadar's parent company Future, heavy use of chatbots such as Gemini and ChatGPT has actually fallen over the past year.

The reasons behind the shift are interesting too: users are mentioning concerns over privacy, a preference for human interaction, and worries about becoming too dependent on the tech when it comes to why they're avoiding AI.

While AI is clearly reshaping our world in significant and myriad ways, there's also now a noticeable trend of anti-AI sentiment that's worth paying attention to — whether it's to do with protests against data center expansion or worries about AI slop content.

These results are based on questionnaires filled out by 1,008 respondents, and they make for interesting reading — especially when you compare them to the statistics from the same Future survey conducted last year. Here's what the latest responses tell us.

Privacy worries

There are many reasons why people aren't so sure of AI at the moment. Topping the list for our respondents are concerns over privacy, which were cited by 32% of those who completed our survey — that's actually unchanged from last year, so no matter which chatbot is being used, many of us are wary about what it's doing with our data.

Second in the list, mentioned by 31% of survey participants, is a preference for human interaction (this is a new response option, so we don't have anything to compare it to from last year). While AI is certainly now more than capable of holding a conversation on just about any topic imaginable, it seems we're still seeking out flesh-and-blood relationships.

In third place we've got those who are happy without AI, which accounts for 29% of people surveyed — that's actually a drop of 18% over the previous survey, perhaps a sign that a significant chunk of people have realized they are happy chatting to AI on a daily basis (it's certainly getting more ubiquitous by the day).

Next on the list are fears around becoming too dependent on the tech, and concerns over AI misrepresenting a 'voice' or personality (in emails, for example), which account for 26% and 24% of respondents respectively. It seems a lot of us are worried we'll get to the stage where we can't do without AI for even basic tasks — not good when there's an outage — and that our communications will turn into generic, bland AI averageness.

Neither of those two responses were available last year, but the next two came in at exactly the same level as in the previous survey: a lack of awareness of AI's capabilities (19% of respondents) and a lack of interest in what AI could do (17% of respondents). It seems the AI hype hasn't quite reached everyone yet.

Rounding out our respondents' list of barriers to AI usage, we've got technical complexity (15%, down 24%), skepticism over whether AI can be helpful (14%, down 21%), environmental and energy impact concerns (11%, a new category), and being philosophically opposed to AI as a concept 10%, down 24%).

Regular AI usage

When answering the question of how often chatbots like Gemini or ChatGPT are used, 17% of those quizzed said "daily or almost daily", while 14% said several times a day. That's nearly 1 in 3 people using generative AI at least once a day or so, though that's actually fewer people than last year — could the novelty of the tech be wearing off?

The 14% figure for the most enthusiastic users has actually dropped by 31% over the course of the past 12 months, so while usage is exploding in areas such as coding, it seems as though the overall trend is in the opposite direction (perhaps influenced by the increasing amount of AI stuffed into our apps).

As for the less frequent users, 21% of respondents said they were using AI chatbots once a week or a few times a week, 11% ranked their usage as being a few times each month, and 8% said they used generative AI less than once a month — apparently dipping in and out for occasional chats.

When it comes to complete refuseniks, 30% of the survey participants said they don't use AI chatbots at all, which is a hefty 24% jump from last year. That suggests more and more people are taking against the very idea of AI, or finding it completely useless for day-to-day life and work.

There is a divide in the respondents worth mentioning between those who read Future publications and those who don't: more than 42% of Future readers use generative AI on a daily basis, with 21% using it multiple times a day, which overall is around double the figure for non-readers.

So that's the state of play for consumers and AI at the moment then, according to the people we surveyed. With generative AI advancements showing no signs of slowing down, it's going to be very interesting to see how these figures might have changed again in another 12 months.

Lately, I've noticed a growing number of celebrities and influencers are talking about AI. Some seem to be partnering with tech companies. Others have positioned themselves as AI evangelists, encouraging their audiences to embrace the technology before they get left behind.

Among those who have generated significant attention are Reese Witherspoon, Mel Robbins, Sandra Bullock and Matthew McConaughey. Their messages do all differ, but they tend to orbit the same idea, which is that AI is here, it's important and you'd better get on board fast.

The thing is, the more a celebrity tells me I need to use AI, the less I want to. And judging by public sentiment, I'm really not alone.

I don't think everyone talking positively about AI is acting in bad faith. Some may genuinely believe it will improve people's lives. Others may have investments, partnerships or financial incentives tied to the industry's success. (And that's hardly unusual — celebrity endorsements have always existed.) Some may simply be repeating the dominant narrative without stopping to consider how influential they are.

But rather than try to understand their personal motivations, what interests me more is the growing gap between the way AI is being promoted and the way many people actually feel about it.

Because while some public figures seem convinced that widespread AI adoption is inevitable, public trust in the technology remains surprisingly low. Survey after survey finds that many people are cautious, sceptical or actively worried about AI. And it's not difficult to understand why.

A growing disconnect

Conversations about AI have moved really quickly over the past year. We now find ourselves debating copyright, creative labor, deepfakes, misinformation, surveillance, environmental costs, job displacement and the growing concern that outsourcing too much thinking to machines may come with cognitive consequences of its own. On the morning I'm writing this, there are fresh headlines about the grim realities of data centre expansion.

At this point, there are so many legitimate concerns surrounding AI that it's difficult to keep track of them all.

Meanwhile, some of the companies that initially appeared determined to replace workers with AI have been rowing back their plans. We've seen reports of AI-generated content requiring extensive human correction, customer service experiments failing to meet expectations and organizations discovering that replacing people is a lot harder than they first imagined.

That's why I find the celebrity enthusiasm so fascinating. Because while some public figures are urging people to embrace AI before they get left behind, many people seem to be moving in the opposite direction.

The comments beneath posts promoting AI are often filled with scepticism. Articles about AI backlash are becoming increasingly common. And when I asked my own social media audience how they felt about celebrity AI endorsements, many expressed similar concerns.

Becky Hughes told me: "All of this makes me more reticent than ever to use social media or adopt new technologies, because the safest option seems to be not to engage at all."

Jay Vera Summer said: "When I see celebrities do things like this, I wish I knew more about their stock portfolio. Especially when it's coming from people who usually don't give financial or career advice."

Whether or not all of our suspicions are fair, I think they suggest something really important is happening here, which is a growing trust gap.

At least from where I’m sitting, it seems people no longer automatically assume that enthusiasm for AI is neutral. They're increasingly wanting to know who benefits from the pro-AI messages, who profits and whose interests are being served when the technology is being promoted so aggressively.

Fear over specificity

If you actually look at what much of the celebrity messaging amounts to, it's surprisingly hollow. Learn AI. Use AI. Don't get left behind. It’s inevitable.

And what fascinates me is how little specificity there tends to be alongside it. What exactly should people be using AI for? Which tools? In what contexts? For what benefit? What are the trade-offs? What are the risks?

I haven't seen many celebrities get into any of that. To be fair, even many of the people building, investing in and advocating for AI rarely spend much time on the details. Instead, the conversation often just gravitates towards fear.

The fear of becoming obsolete. The fear of missing out. The fear of being left behind by a future that everyone else supposedly understands. As someone who has spent years covering technology, that kind of rhetoric always makes me uneasy.

And that's not because I think AI won't have a place in the future. I think it almost certainly will, for better and for worse. But because "you'll get left behind" isn't really an argument. It's an appeal to our anxieties so that you’ll act fast without thinking. And it’s encouraging adoption without fully engaging with the reasons people might be hesitant in the first place.

AI as a feminist issue

I find this particularly interesting when AI is framed as a feminist issue. Earlier this year, The Cut described this phenomenon as the "girlbossification" of AI, giving a name to the growing trend of influential women encouraging other women to embrace the technology or risk falling behind.

Several prominent women have made versions of this argument. And they’re sort of right. In some studies, women have adopted generative AI more slowly than men. But the gap appears to be driven partly by risk, ethics, and workplace factors, not just technical ability. And women have plenty of reason to be concerned about the risks.

We know that women and girls have been disproportionately affected by some of AI’s most disturbing uses, including deepfake pornography, AI-generated image abuse, and sextortion. In one UN estimate, up to 95% of online deepfakes are non-consensual pornographic images and 99% of those targeted are women. I know this isn't theoretical because I've experienced a version of it myself.

Against that backdrop, telling women they simply need to embrace AI can feel completely disconnected from reality. It risks treating healthy scepticism as ignorance when, in many cases, it seems to me that it’s a response to genuine concerns and lived experience.

The recent partnership between Kylie Jenner and Meta feels particularly relevant here. The campaign positions AI-powered glasses as fashionable, desirable and aspirational. And in some ways that's exactly what celebrity endorsements have always done, take a technology and make it feel culturally normal.

But that's exactly why these messages deserve scrutiny. At the same time women are being encouraged to embrace AI-powered devices, there have already been multiple stories of women being unknowingly recorded by smart glasses. Which to me highlights the very real concerns around privacy, consent and surveillance that often get overlooked in conversations about the latest cool new tech on the block.

More informed than you think

I've seen some people online brush all of this conversation off and argue that we shouldn't be taking celebrities seriously anyway. But like it or not, they do help shape public narratives. They influence what people pay attention to, which questions get asked and which concerns get ignored.

And right now, many of those narratives seem to be built around this strange assumption that resistance to AI exists because people don't understand it. Well, I've spent the past year talking to people about AI, and I suspect the opposite is often true.

Many people understand enough to have concerns. They've tried the tools. They've seen both the benefits and the downsides. They're making conscious decisions about how much of their work, creativity, relationships and attention they're willing to hand over to AI systems.

That's why I find so much of the celebrity messaging unconvincing. The more people tell me I have to use AI, the more I want to pause and ask: okay why? And I know I'm not alone.

And that’s not because people are afraid of the technology. If anything, that framing completely misses the point. What I see instead is caution, scepticism and a willingness to actually ask the difficult questions about where this technology is taking us.

Because I think whenever someone insists a certain future is inevitable, our alarm bells should start ringing. That's when we need to ask: okay, whose version of the future are you trying to sell?

• OpenAI has brought big changes to ChatGPT’s desktop app
• ChatGPT and Codex have been merged into a single app
• Users are unhappy with many of the new updates

OpenAI has been running its Codex vibe coding platform as a separate tool from the regular ChatGPT app for a while now, but that’s all changed with the artificial intelligence app’s latest update. The two products have been merged into one ‘super app’ — but many users are deeply disappointed with the move.

If you download ChatGPT today — or update an existing installation — you’ll find an app that looks very different to the one you’re used to, which could be particularly disorienting if you’ve used the artificial intelligence (AI) chatbot for some time now.

For example, you’ll now find a new toggle in the top-left corner of the app that lets you switch between ‘ChatGPT Work’ and ‘ChatGPT Codex.’ Instead of containing buttons for your library, projects, and apps, the left-hand sidebar now houses scheduled tasks and plugins. Projects and tasks sit below in their own sections.

What used to be the ChatGPT app still exists, but has been renamed to ChatGPT Classic. The standalone Codex app, meanwhile, has become the main ChatGPT app. Both ChatGPT Work and ChatGPT Codex share plugins, and while both are capable of similar tasks, Codex shows more of the technical details that Work obscures in a bid to be more user-friendly.

An unpopular move

Unfortunately for OpenAI, this major change has not gone down too well. Apple pundit John Gruber, for example — long a fan of OpenAI’s decision to make ChatGPT a native Mac app that conformed to macOS conventions — described the change as “the day OpenAI f*cked up the ChatGPT Mac app.” He also pointed out that while the old ChatGPT app weighed in at a lightweight 159MB, the new edition has ballooned to 1.5GB.

I’m certainly not a fan of the new-look ChatGPT either. I have little interest in Codex and use ChatGPT to work through problems, often revisiting old conversations to add new details. Yet in the new app, your recent conversations are hidden and require many more clicks to access. Only your five most recent threads are shown, and to access more you have to click Chat > See All to uncover them. Previously, they were right there waiting for you in the sidebar.

As well as that, my existing projects seem to have totally disappeared. I only had a couple, but they were related to a small-claims case I was involved in, so seeing all my relevant threads scattered and disorganized in a long list of chats is a major pain. For anyone juggling a plethora of projects, this change could be disastrous. And what’s more, Custom GPTs also seem to have vanished into the ether.

I know I’m not alone in my disappointment. Over on Reddit, OpenAI’s decision has not been well received, with user powprodukt summing up the mood by saying: “ChatGPT projects and custom GPTs are the heart of why I use that app. Without these features there is no feature parity.”

Other users took issue with the new user interface, with Kaotic987 exclaiming “Christ almighty, this is so bad!” Eriane, meanwhile, cynically suggested that OpenAI “vibe coded everything without testing it.”

Whether you even get ChatGPT Classic seems up for debate. It’s available for Windows in the Microsoft Store and can be downloaded separately from the new ChatGPT app. On macOS, however, I updated to the new edition and ChatGPT Classic is nowhere to be seen. Luckily, I backed up the old app before updating, so I can still access it, which might be the way to go if you’re a Mac user and don’t want to lose all your projects and custom GPTs.

The move to combine ChatGPT and Codex into a form of ‘super app’ has already alienated plenty of users, myself included. From the questionable user interface to the absence of projects and custom GPTs, there’s a lot for OpenAI to work on before it can get the app back to a good place.

• Report claims China's AI consumption is now over 100 trillion tokens per day, up from 100 billion in justtwo years
• Enterprise customers are increasingly paying per usage, not per seat
• General consumers will likely keep their fixed, monthly pricing

AI token consumption in China has increased "more than 1,000-fold" in around two years, with National Bureau of Statistics data claiming daily token consumption rose from 100 billion at the start of 2024 to 100 trillion by the end of 2025.

By March 2026, China was consuming around 140 trillion tokens per day, which SCMP illustrates as 100,000 tokens per person per day if every single Chinese citizen, including children, pensioners and non-workers, were to use AI.

"Today, the token economy is emerging," Yin Hao of the Chinese Academy of Sciences said in a recent Beijing conference, the South China Morning Post reported.

AI consumption-based pricing could replace per-seat pricing

With users consuming more tokens than ever before, AI providers are now under pressure to cover the growing costs. Industry experts like Hao now believe that token-based pricing could soon replace traditional subscription models.

This could see subscribers paying a fee based on how much AI they actually use, rather than paying a fixed monthly charge per user.

AI providers globally are also increasingly recognizing growing token consumption and are adapting their pricing strategies to match, including OpenAI and Anthropic. Others, like Zendesk, are charging per meaningful outcome rather than per seat or per token.

However, for now, these pricing shifts are primarily affecting enterprise customers consuming huge amounts of tokens. General consumers logging into their preferred chatbot, like ChatGPT or Gemini, are unlikely to be affected as compute becomes cheaper.

Mistral Vibe is one of the more interesting AI platforms to come out of Europe. The product, formerly called Le Chat, was rebranded by Paris-based Mistral AI in May 2026 to mark its shift from a conversational assistant to a full agentic platform covering chat, work automation, and cloud-based coding. It's a meaningful change, not just a naming exercise.

Two new operating modes, Work and Code, are the engine behind that ambition. Work Mode runs multi-step tasks across connected business tools. Code Mode handles remote coding sessions inside isolated sandboxes and delivers work through to a pull request. Together they put Vibe in more direct competition with ChatGPT and Claude than its Le Chat days suggested.

At TechRadar Pro, we've been reviewing business software since 2012 and our AI coverage has become some of our most-read work. That includes our AI tool roundup, our 2026 vibe coding buying guide, and deep dives on platforms like OpenClaw or Moltbook.

What is Mistral Vibe?

Mistral Vibe is a chat, work automation, and coding platform developed by Mistral AI. It's available via web browser and mobile apps on iOS and Android, and runs on Mistral's own model family, from the lightweight Small 3.1 to the flagship Large 3 and the reasoning-focused Magistral line.

The platform launched in February 2024 as Le Chat, originally a general-purpose assistant. Over the following two years, Mistral layered in web search, voice mode via its Voxtral audio model, a Canvas document editor, image generation through Black Forest Labs Flux Ultra, persistent memory, and project folders.

The May 2026 rebrand signals that Mistral sees Vibe as an enterprise-grade product. Work Mode and Code Mode are aimed at professionals who need an AI that can execute multi-step tasks across connected tools, not just hold a conversation.

Mistral Vibe: At a glance

Buy it if…

• You handle sensitive client or business data. No Telemetry Mode gives Pro subscribers a contractual guarantee that nothing they type is used to train Mistral's models. That level of assurance at $14.99 per month is unusual among major AI chat platforms.
• You want agentic features without enterprise pricing. Work Mode handles multi-step tasks across Gmail, Slack, Notion, and other connected tools. Code Mode manages full coding sessions through to a pull request. Both are available on the Pro plan.
• You're a developer watching API costs. Mistral Large 3 at $6 per million output tokens significantly undercuts GPT-5.4 ($15/M) and Claude Sonnet ($15/M) at the flagship tier, and the API bills only for tokens used.

Don't buy it if…

• The free plan is your entry point for serious evaluation. At around 25 messages per day with no canvas and no remote coding access, the free tier doesn't give you a fair picture of what the platform can do.
• Your team is embedded in Microsoft 365 or Google Workspace. Vibe connects to business apps via MCP, but native integration with the major office suites is not as developed as what you'd get from Copilot or Gemini for Workspace.
• You're assessing the Team plan for a small group. The jump from Pro ($14.99) to Team ($24.99/user/month) is steep, and the main additions (admin controls and more storage) may not justify the cost for teams of fewer than ten people.

My time with Mistral Vibe

I tested Vibe across its Pro plan over several weeks, using it for research tasks, document drafting, and code work. The chat interface is clean and approachable. Anyone familiar with ChatGPT or Claude will navigate it without confusion. What caught me off guard was how efficiently Work Mode handled complex, multi-step research requests, pulling from connected sources and drafting a structured Canvas output in a single run.

Code Mode held up well for the tasks I threw at it. I ran a session to scaffold a simple API integration, and the agent handled writing, testing, and preparing a draft PR inside the sandboxed environment. I stepped in twice to give it additional direction, but that's consistent with what you'd expect from any AI coding agent at this stage.

The No Telemetry Mode stood out as a differentiator. Enabling it took seconds and gave me real confidence when working with business-related documents. That kind of data control is typically reserved for enterprise tiers at other major platforms, so finding it on a $14.99 plan is a real differentiator.

Mistral Vibe: Features

Vibe's feature set has grown considerably since the Le Chat launch in 2024. Alongside standard chat, the platform now covers web search, image generation, voice input, a Canvas document editor, Deep Research, memory, and project organization tools. Pro subscribers get access to Work Mode and Code Mode, which is where the real differentiation lies.

Work Mode is the headline addition from 2026. It turns Vibe into an execution agent that can read emails and calendars, draft documents in Canvas, run recurring scheduled tasks, and push outputs to Notion, SharePoint, or Slack. Every step is visible in the interface. The platform asks for explicit approval before any action that modifies data or sends a message.

Code Mode targets developers specifically. It connects to GitHub, GitLab, Jira, and Linear, running sessions in an isolated sandbox. A /teleport command lets you move a session between the Vibe web app and a local terminal without losing context. Parallel sessions are supported on Pro and above, which is useful for running background jobs while staying in another workflow.

Memory, currently in beta, lets Vibe store your preferences and recurring context across conversations. You can view, edit, or delete these entries at any time, and turning the feature off is a single toggle in confidentiality settings. Image generation via Black Forest Labs Flux Ultra is solid for a chat platform.

The main gap in the feature set is transparency around usage limits. The free tier runs at lower quotas than Pro across searches, image generations, and messages, but Mistral doesn't publish the exact numbers on its pricing page. That makes the free-to-Pro comparison harder to assess than it should be.

Overall, Vibe covers more professional use cases than most platforms at this price point. The area where competitors like Copilot and Gemini maintain a clear lead is native productivity suite integration, which Mistral has not yet matched.

Mistral Vibe: User experience

The layout is straightforward: a left sidebar for conversation history and project folders, a central chat window, and model or tool selectors accessible from the input bar. Switching between chat, Work Mode, and Code Mode happens within the same interface rather than routing you to a separate product URL.

The mobile apps mirror the web experience closely. Search, canvas, image generation, and voice input all carry over, which isn't a given with AI chat platforms on mobile. Onboarding is minimal, which suits experienced AI users but may leave newcomers without much guidance on how to get started with the more complex agentic features.

Mistral Vibe: Customer support

Free and Pro users access support through the help center widget on the Mistral site. Response times are not published, and there's no live chat or phone support at these tiers. The help documentation covers the most common issues in reasonable depth, but for billing questions or edge-case technical problems, you're relying on ticket-based email support.

Enterprise customers get a dedicated support workflow with priority routing. Requests go through the same widget but are flagged and handled separately based on account type. For teams in finance, healthcare, or other regulated sectors where response time matters, that distinction is a real consideration when deciding between Team and Enterprise.

Mistral Vibe: Pricing

• Free plan: Around 25 messages per day with limited web searches, reduced image generation, no canvas creation, and no remote coding.
• Pro at $14.99/month: No Telemetry Mode, full canvas access, Work and Code Mode, 5x more web searches, more image generations, and pay-as-you-go Vibe coding beyond included limits.
• Team at $24.99/user/month (or $19.99 on annual billing): Shared workspaces, admin controls, domain verification, data export, and higher storage limits.

The free plan gives you a taste of chat quality and basic search, but the restrictions mean you won't get a representative experience. Pro at $14.99 is, by most comparisons, the cheapest premium AI chat subscription from any major provider. The No Telemetry Mode and agentic modes are hard to find elsewhere at this cost.

The Team plan's value depends on your use case. For small teams, the mainly administrative additions over Pro may not justify the per-user cost. Enterprise pricing is negotiated directly with Mistral and covers SAML SSO, on-premise deployment, custom model training, and dedicated support. The API runs on a fully separate billing track with no monthly minimums, making it accessible for developers at any scale.

Mistral Vibe alternatives you should consider

• ChatGPT (OpenAI): The most established option for general productivity, with deeper Microsoft 365 integration and a broader ecosystem of plugins. It costs more at comparable tiers but suits teams already embedded in the Microsoft environment.
• Claude (Anthropic): A strong choice for long-document analysis and nuanced writing. Pricing is competitive and reasoning quality is high, though data privacy controls at the consumer tier are less explicit than Vibe's No Telemetry Mode.
• Gemini for Google Workspace: The best fit for teams already using Google's suite. Native Calendar, Docs, and Gmail integration outpaces what Mistral currently offers for Google-centric workflows.

How I tested Mistral Vibe

• Ran prompts across standard chat, Web Search, and Deep Research modes to evaluate response accuracy, source quality, and multi-step research handling across a range of topics and document types.
• Tested agentic task execution through Work Mode, including document drafting via Canvas and recurring task scheduling, and ran coding sessions through Code Mode against a GitHub-connected project to assess end-to-end agent performance.
• Verified plan details and API rates against Mistral's official pricing page, cross-referenced them with third-party pricing analyses where official documentation was vague, and assessed support options through the Mistral Help Center.

I tested Mistral Vibe on its Pro plan over several weeks using a mix of daily productivity tasks and structured feature evaluations. Pricing data was verified against the official mistral.ai pricing page, with third-party sources used for cross-reference on API rates and plan limits where Mistral's documentation was unclear. Support quality was assessed through available help center documentation and publicly reported user experiences.

HIX.ai started as an AI writing assistant and has since expanded into something considerably more ambitious. The platform now markets itself as an "ultimate AI agent workspace," bundling tools for content writing, deep research, image and video generation, slides creation, coding, and multi-model chat under a single interface. It's a notable evolution, even if the transition has introduced some rough edges.

What sets HIX.ai apart from simpler AI chat tools is the sheer breadth of models on offer. Users can switch between GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, DeepSeek-V4, and others from within the same session. For teams or individuals who want to compare outputs or pick the right model for a specific task, that flexibility has real practical value.

We've been reviewing business software at TechRadar Pro since 2012 and AI platforms have become one of our most-covered categories. In recent months, we've also published our best AI tools roundup and deep-dives into platforms like OpenClaw and Moltbook. HIX.ai falls squarely in our coverage area.

What is HIX.ai?

HIX.ai is a web-based AI platform that consolidates multiple AI capabilities into one workspace. Its agents cover text generation, long-form article writing, email composition, image creation, video production, presentation slides, deep research, and coding assistance. You access everything through a shared interface rather than juggling separate subscriptions for each function.

The platform targets a wide audience: content marketers who need to produce copy at scale, students looking for a homework or essay tool, small business owners who want a single AI tool to handle email, social posts, and research, and knowledge workers who want to stay inside one tab instead of bouncing between apps.

HIX.ai also offers a browser extension that works across Gmail, Google Docs, LinkedIn, Twitter, and most major web platforms. A desktop app is available alongside the main web interface, making it one of the more cross-platform options in this category.

HIX.ai: At a glance

Buy it if…

• You want one platform for everything. HIX.ai removes the need for separate subscriptions for writing, image generation, and research. That consolidation saves both time and money for users currently paying for multiple tools.
• You want model choice without switching apps. The ability to call GPT-5.5, Claude, Gemini, and DeepSeek from a single interface is a real convenience, especially when different tasks call for different models.

Don't buy it if…

• Billing transparency matters to you. Multiple users across review platforms have reported being charged for annual plans they believed were monthly. HIX.ai's refund window is three days, which leaves little room for course correction.
• You need predictable usage. Credits don't roll over between billing cycles, and the "Unlimited" label on paid plans applies only to standard credits. Advanced features are still capped, which surprises users expecting truly open access.

My time with HIX.ai

I spent several sessions testing HIX.ai across its main agent categories. The multi-model chat interface impressed most immediately: switching from Gemini 3.5 Flash to Claude Opus 4.8 mid-conversation takes a few clicks, and responses feel snappy enough for practical work. The deep research agent was particularly useful, pulling together sourced summaries on a complex topic faster than a manual web search.

The browser extension held up well in Gmail and Google Docs. Highlighting a paragraph and getting a rewrite option surfacing instantly, without opening a new tab, is the kind of friction reduction that matters when you're working fast. The image generation tools produced serviceable results, though I found the credit consumption less predictable than I'd have liked.

Where the experience wobbles is in understanding what you're paying for. The credit system splits between standard and advanced buckets, and it isn't always clear which agent or model draws from which pool. New users should read the plan details carefully before committing to a billing cycle.

HIX.ai: Features

HIX.ai's feature roster is unusually wide. The platform covers AI chat across multiple frontier models, a long-form article writer, email and social copy tools, a paraphrasing and summarization engine, image generation via Midjourney, Flux, and GPT Image 2, video generation via Seedance 2.0, Veo 3.1, Sora 2, and others, a slides agent, and a deep research tool that produces structured reports with real-time internet access.

The deep research agent is one of the more differentiated offerings. Rather than a standard web search, it compiles structured, sourced reports on a given topic, making it useful for due diligence, competitive analysis, or academic research. Real-time internet access is included on paid plans, which keeps output current rather than frozen at a training cutoff.

On the writing side, the 120+ AI writing templates cover a broad range of formats, from ad copy to blog outlines to product descriptions. The long-form article writer handles structure and headings reasonably well, and the AI Writer agent supports markdown and rich text outputs, which makes it easier to paste results into a CMS or document editor.

The platform's weakest spot by reputation is HIX Bypass, its AI humanizer tool. Multiple independent reviews and user reports flag inconsistent results against modern AI detectors, particularly after Turnitin's 2025 updates. HIX.ai positions this as a key feature, but anyone relying on it for academic or professional compliance should test it carefully before committing.

The multi-agent coordination feature, available across all tiers, lets different agents pass context between tasks. In practice, a research output can feed directly into a writing task without manual copy-paste, which speeds up structured workflows considerably.

The video and image agents round out a platform that few competitors match in raw breadth. Whether that breadth delivers depth at every point is a fair question; some agents feel more developed than others. But for users who want to consolidate tools, the coverage is hard to beat.

HIX.ai: User experience

The interface is clean and navigation between agents is straightforward. The sidebar layout keeps the main workspace uncluttered, and the model switcher is tucked sensibly into the chat header rather than buried in settings. For new users, the onboarding path is light: a working chat session is within reach in under a minute.

The browser extension is where the experience stands out. HIX.ai has put real thought into how it surfaces inside other platforms: the quick-action bar appears when you highlight text, and the full sidebar opens on demand. For anyone doing a lot of writing inside web apps, this integration removes the switching cost that blunts many AI tools, and the learning curve for core features is minimal.

HIX.ai: Customer support

Support is available via email across all plans, with priority support reserved for paid subscribers. The HIX.AI Community forum is open to all tiers and covers common issues and feature requests. Response times on the email channel have drawn complaints in user reviews, with some reporting delays of multiple days for billing queries specifically.

The three-day refund window is short by industry standards. Platforms like ChatGPT Plus and Jasper offer more flexible cancellation terms. Given the annual billing complaints that appear consistently in user feedback, the lack of a longer trial-to-refund period is a genuine sticking point for new subscribers.

HIX.ai: Pricing

• Free plan: $0/month with 20 credits, enough to test core features but limited for sustained work.
• HIX AI Max: Billed annually; includes unlimited standard chat models, advanced models, video, image, and all agent features including deep research, AI slides, and priority support.
• HIX AI Pro: Billed annually at a higher tier, with different credit allocations across advanced features.

The free plan offers real access to the platform with no credit card required, though 20 credits runs out quickly once you start testing agents. Paid plans are priced on a credit system that splits standard and advanced usage into separate buckets. SoftwareSuggest, which updated its pricing data in April 2026, lists the starting price at $13/month on annual billing, with higher tiers scaling up from there.

HIX.ai's various product verticals (AI Writer, HIX Bypass, the browser extension, EssayGPT) each carry separate subscription structures for legacy users, adding complexity to an already layered pricing model. The newer unified workspace tiers (Free, Max, and Pro) simplify this somewhat, but the credit system still rewards careful reading before you subscribe. Annual billing delivers the best per-month rate, though it comes with the billing caveats noted above, and there is no public API for developers.

HIX.ai: alternatives you should consider

• ChatGPT Plus ($20/month): Similar multi-model access via GPT-5 and o3, with simpler pricing and a stronger track record on billing transparency.
• Jasper ($39/month billed annually): Focused on marketing and long-form content. Less breadth than HIX.ai but stronger output consistency on brand voice tasks.
• Perplexity Pro ($20/month): If deep research is your primary use case, Perplexity's research-first approach and real-time citations are hard to beat at this price point.

How I tested HIX.ai

• Ran multi-agent workflows across the chat, deep research, and writing agents to evaluate inter-agent handoffs and model-switching convenience.
• Tested the browser extension inside Gmail and Google Docs on multiple sessions, focusing on responsiveness and quality of inline suggestions.
• Evaluated output quality across standard chat, long-form article generation, and the image creation tools using consistent prompts to benchmark against competitor platforms.

I spent approximately two weekdays with HIX.ai across different workloads, including research tasks, marketing copy drafts, and content editing sessions. Pricing information was cross-referenced against the official HIX.ai pricing page and third-party review sources updated in 2026. Feature details were drawn from the platform's live interface and official documentation.

• Palo Alto Networks CEO Nikesh Arora says 90% reduction in AI pricing is needed to drive widespread adoption, 20% reduction by next year
• OpenAI's latest GPT-5.6 Sol model is 54% more efficient across agentic coding tasks
• Shift to consumption-based models is making high costs even more pronounced

Palo Alto Networks CEO Nikesh Arora has warned that AI pricing is still too high for widespread enterprise adoption, arguing that it should be drastically cheaper.

Although model efficiency and performance have come a long way in recent years, he believes cost remains one of the biggest barriers for widespread adoption.

In a recent interview with CNBC, Arora laid out how he believes token prices need to fall by as much as 90%, however he acknowledges that the change is unlikely to happen immediately, instead welcoming a 20% price reduction over the next year or so.

Is AI pricing holding enterprises back from widespread adoption?

Arora's comments come in response to OpenAI's latest announcement, revealing that its latest GPT-5.6 Sol model is now 54% mode token-efficient across agentic coding tasks – a major improvement.

Though Palo Alto Networks' CEO admitted that this in itself is a good start, it's not enough of an improvement to drastically change pricing for major enterprises at the moment.

"It's important to understand the demand continues to be infinite," he added, implying that cost remains prohibitive. The issue of cost is also becoming more of an issue as AI companies evolve their pricing strategies, with customers moving toward consumption-based models rather than flat monthly fees per user.

Importantly, Arora's vision for cheaper AI isn't unfounded. Virtually all previous technological revolutions have already followed a similar pattern. In the case of AI, compute would become cheaper, models would become more efficient and competition could keep prices low, all working in the favor of enterprise customers.

• OpenAI has introduced new agentic ChatGPT Work tools for all users
• The tool is powered by the latest GPT-5.6 model and Codex
• Could this be the app's unofficial evolution into a superapp?

OpenAI has lifted the wraps off a new agentic tool within ChatGPT to help workers "take on more ambitious tasks," covering workflows like gathering information, creating finished materials and handling long-running autonomous tasks.

The new tool, called ChatGPT Work, marks a shift in how knowledge workers are set to use AI, evolving from generative AI to autonomous agentic AI.

ChatGPT Work, largely seen as an interface for knowledge workers, brings together tools like the company's latest frontier model GPT-5.6 and Codex.

ChatGPT Work is an autonomous knowledge worker's dream

In an announcement, the company noted how Codex was born out of the need for coders and developers to get access to agentic capabilities, but now, more than a million non-coding workers use it for other tasks.

The company also declared that nearly all of its internal teams use Codex and ChatGPT Work to speed up their workflows.

All users running the app on their macOS or Windows desktop can get access to ChatGPT Work for free, including non-paying subscribers. Pro, Enterprise and Edu plans will get mobile and web access first, but Plus and Business plans will get access within days.

Further aiding ChatGPT Work's autonomy, the tool also works with third-party plugins via the '@' command to obtain further context and access to working files.

More broadly, as the app evolves into a superapp to combine all of OpenAI's tools in one workspace, desktop users can now grant the client permission to control the computer on their behalf, similar to Anthropic's Claude Code.

"This is the first step towards a broader vision for ChatGPT," the company summarized, "where intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality."

• The Mac mini has emerged as an affordable system for agentic workloads
• Apple has seen "incredible demand" for the Mac mini and Mac Studio
• Apple silicon can handle an agentic AI while other architectures use a GPU and CPU

If you’re looking for the best way to explore and deploy agentic AI without breaking the budget, the Mac mini might be just what you’re looking for.

Apple’s Doug Brooks has expressed enthusiasm for how the Mac mini and Mac Studio desktop computers are capable of handling agentic AI tasks, thanks to Apple silicon, the ARM-based SoC that the company has introduced over the past half decade.

Success with local AI on these machines has been attributed to design choices made before the arrival of advanced LLMs, with the evolution of Apple’s Neural Engine highlighted as a key factor.

How the Mac mini is ideal for agentic AI

Brooks is the senior product manager of Apple silicon, and referred to the “incredible demand” for Mac minis and Mac Studios when speaking to The Deep View before WWDC 2026.

Describing the Mac mini as an “amazing system” that can “tap into the strengths of Apple silicon and unified memory in a very power-efficient way, and increasingly they're delivering compelling price-performance as well.”

The price point of a Mac mini – compared to the more expensive Mac Studio – makes it particularly suited to teams exploring agentic AI but without the budget to pay for tokens and larger systems.

Neural Engine technology dates back to the A11 chip, and its evolution and inclusion within the current generation of Apple chips, and its high-performance, power-efficient compute processes are pivotal in delivering machine learning to the desktop.

As many AI tools were available first on the Mac (or released exclusively for macOS), it seems that upgrading to the latest Mac mini or switching from Windows has been instrumental in demand.

Mac mini: amazing for AI

Apple’s work on AI has seen deployment in everyday use across computers, tablets, and smartphones, and the company has been a leading exponent of hybrid AI, where an agent can “decide what needs to happen locally and what needs to happen in the cloud based on the workload.”

“For agentic workloads, people often want a system that's under their control, isolated from their primary machine, and capable of running 24 hours a day, seven days a week.”

But it is the strength of the Apple Mac mini and Apple Studio – as well Apple’s notebooks – in handling AI that seems to have enthused Brooks the most. He cites security and economics as concerns for developers and creators who are now realising that they can handle AI workloads sitting at their desk – whether using a Mac mini or something more powerful.

• Bosses are seemingly confused that mass AI in favor of human workers isn't always a success
• AI operating costs are often higher than expected
• However some firms are happy to adapt and refocus where needed

Bosses are being left confused at the high cost of moving to AI-centric models, with many seemingly baffled that replacing human workers with agents isn't instantly saving them huge amounts of money, new research has claimed.

A new KPMG report found nearly a third of business leaders reported some difficulties with getting to grips with AI operating costs in their organizations.

The news comes as several major AI providers, including the likes of Anthropic and OpenAI, have moved some services toward usage-based billing, rather than flat-rate subscriptions, in recent months.

Making AI effective

The report saw KPMG survey 2,145 senior leaders across 20 countries, finding 29% were struggling to understand the rise in operating costs as they looked to scale AI across their business, with a similar proportion also highlighting a limited understanding of AI costs and economics as a major challenge to deploying AI agents.

"As usage-based pricing models become more common, many organizations are still building the capabilities required to forecast, monitor, and manage AI spending effectively," KPMG said.

When things do go wrong, the report highlighted how leaders were often unclear who should take responsibility, particularly in the case of hallucinations or errors by AI models.

It noted that why having human leadership be accountable is important, "governance ultimately succeeds or fails through day-to-day operating practices."

"Organizations need clear rules for when employees can intervene, who owns AI-related costs, how AI outputs are reviewed and what happens when systems fail. While most organizations report having at least some governance mechanisms in place, relatively few describe these practices as fully embedded," the report said.

When costs have outweighed the expected value, the report found a surprising amount of contrition from its participants, with nearly half of organizations saying they had rephased AI deployments in that case.

"These actions do not signal reduced confidence in AI," the report warned. "Rather, they suggest a growing willingness to evaluate where AI creates meaningful value and where it does not. Organizations appear increasingly focused on concentrating investment where expected returns are strongest."

"We’re seeing a clear divide between organizations with leadership accountability at the top and those without," added Steve Chase, KPMG Global Head of AI and Digital Innovation, KPMG International.

"These companies are seeing materially better results across the board such as greater confidence, higher value realization and established ROI."

• OpenAI plans to hire a $200k Investment Banking expert as it continues to target finance
• The company wants to improve financial accuracy, consistency and overall quality
• ChatGPT already has its own personal finance tool

OpenAI is advertising a new "Subject Matter Expert, Investment Banking" role within its team, serving as the latest hint that the ChatGPT maker wants to expand its reach into the finance sector.

This latest position strongly implies that the company wants to increase more sophisticated Wall Street type tasks, rather than be an all-purpose chatbot, and it makes sense.

We've already seen OpenAI and rivals like Anthropic heavily target certain sectors – banking and law being two of the most evident in recent months.

Could OpenAI be expanding its reach into finance?

OpenAI described investment banking as one of the most demanding forms of knowledge work, putting pressure on workers to "synthesize fragmented information, exercise judgment under pressure, and produce precise, defensible models, analyses and client materials."

Being that the job was posted in an entire blog post-style announcement rather than on a dedicated careers platform, it's clear the role will be a high-level one reporting to senior leaders and will play a considerable role in the company's direction. The salary of $185,000 to $205,000, plus equity, is also indicative of the role's importance.

The announcement highlights the importance of quality over quantity, stressing the need for AI to help produce work that's financially correct, traceable and consistent.

OpenAI also wants prospective candidates to help identify the highest-value AI opportunities in investment banking – another clear signal that the company is going after the sector big-time.

With its existing expertise, OpenAI has already targeted the consumer end of the sector with a new Personal Finance tool within ChatGPT that connects to bank accounts to give users insights into their spending habits, investments and savings.

Do you remember the metaverse? If you don’t, don’t worry. In 2026, four years into the AI revolution that’s changing the world forever, you could easily be forgiven for thinking it was a strange fever dream you had back in 2021. You might even have odd memories of seeing a blocky version of Mark Zuckerberg floating about in a Minecraft-inspired hellscape, conducting meetings with people who could walk with no legs, while the real Mark Zuckerberg was looking at the whole thing through VR goggles. At least, that’s my memory of it.

I could be suffering from the Mandela effect, but I distinctly remember something off about the legs. They fixed that in a later version, but that’s my overriding memory — no legs. Oh, and Mark Zuckerberg assuring us that this was the future. He’d spent tens of billions of dollars on it, even changing the company name from Facebook to Meta, just to let us know he was really serious about the metaverse. Even if the legs didn’t work.

The problem was, it looked laughable. While everything in the technology world to do with games and special effects was moving in the direction of hyper-realism, the metaverse was moving in the opposite direction, towards the sort of blocky graphics that small children enjoy. But even that didn’t really answer the most basic question about the metaverse. Why? What possible advantage was there for us all to meet in a VR space where clunky avatars of ourselves could interact… badly?

Then AI happened and Meta abruptly forgot about the metaverse and pivoted towards the mission of putting personal superintelligence in all our hands instead, which sounds as terrifying and dangerous as it actually is, but we are where we are. At least when Zuckerberg was obsessed with the metaverse, we could ignore him. It existed somewhere “over there”, in Meta-land, where we could let him get on with it. Now he’s right up in our business again.

The rise of a bad idea

I’ve always been fascinated by how Zuckerberg got into the metaverse and why he became so obsessed with it. The origins of the metaverse go way back. In March 2014, Facebook bought Oculus, the VR company, for about $2 billion. This was where his passion for VR started. Think of it as the seed, not the full obsession.

By July 2021, Zuckerberg gave a long interview to Casey Newton at The Verge about Facebook becoming a “metaverse company” and described it as an “embodied internet”. Then, on October 28, 2021, his obsession became the company identity. Zuckerberg announced that Facebook the company was becoming Meta at Connect 2021, saying the new company brand would focus on bringing the metaverse to life.

I can see the logic. As a business strategy, that made sense. As a product ordinary people were expected to use, it was much harder to understand. Zuckerberg did not want Meta/Facebook to be trapped inside someone else’s platform again. Facebook had won on social, but on mobile it remained dependent on Apple and Google for distribution, privacy rules, app-store policies and hardware. The metaverse looked like a chance to own the next operating system of social life: hardware, avatars, identity, payments, meetings, gaming, work, commerce — the whole stack. In his 2021 founder’s letter, he framed the metaverse as the “next chapter of the internet” and said Meta would become “metaverse-first, not Facebook-first.”

Facebook had obviously had its problems — it was scandal-ridden. It had let advertisers target vulnerable teenagers, helped fake news spread, and enabled the spread of hate speech linked to atrocities in Myanmar. Perhaps Zuckerberg was looking for a way out of Facebook, and the metaverse offered that.

What I still didn’t understand was why he didn’t see what the rest of us saw — that it looked terrible and offered no real benefit to users. Then I read Sarah Wynn-Williams’ tell-all book about Facebook, Careless People, and it all started to make sense.

To say the book made my jaw hit the floor on several occasions would be an understatement. It’s an absolute page-turner, and your reactions grow from mild amusement to shock, then disbelief, then absolute outrage the further through the book you get. I’m aware of the criticisms of Wynn-Williams: that it is a book written by a disgruntled employee, and that she dodges a lot of personal responsibility for her part in the various misdeeds of the company. However, in another perfect example of the Streisand effect, the fact that Meta obtained a legal order in the United States to prevent her from saying anything negative about the company — at all — made me want to pick it up, and I’m glad I did.

Because now I get it — Zuckerberg seems to have spent years in an environment where too few people were willing to tell him when his ideas weren’t good. According to Wynn-Williams, he was surrounded by sycophants. When he had bad ideas, like the ill-fated Internet.org, he wouldn’t let them go and persisted with them, even when they were obviously going to fail. The people around him enabled him because he was simply too powerful. They even let him win at the board games he liked to play with them at his house or on his jet, and — crucially — he didn’t notice that they were letting him win. I can imagine that in that environment, nobody inside Meta would want to tell Zuckerberg that his metaverse was the equivalent of the emperor’s new clothes, especially if they wouldn’t even risk beating him at Settlers of Catan.

Wynn-Williams only mentions the metaverse in her epilogue. It happened after she was brutally fired from Facebook. Perhaps selfishly, I wish she’d been there for the metaverse period, because I would love to read firsthand accounts of how and why Zuckerberg persisted with such an obviously bad idea.

The fall of the metaverse

Maybe I’m being too harsh on Zuckerberg. The metaverse graphics did get better over time and Apple ventured slightly into the same territory with its Apple Vision Pro, even after the metaverse had turned into a smoldering wasteland. The fact is, people don’t enjoy wearing VR goggles for extended periods of time, and for normal people, VR lacks that one killer app. There doesn’t seem to be anything you can do in a VR space that you can’t do elsewhere much more easily.

The metaverse didn’t really die with a bang, but with a whimper. It faded through layoffs, spending cuts and the AI pivot. If I had to put a date on it, I’d say early 2023 was when Meta’s narrative moved on. In February and March 2023, Zuckerberg started talking about Meta’s “year of efficiency” and announced huge layoffs and cost-cutting. OpenAI had launched ChatGPT in November 2022, and by early 2023, generative AI had swallowed the oxygen that ideas like the metaverse need to survive. Every tech company was talking about AI now, not virtual offices and avatar legs.

The metaverse was over. We all forgot about it and moved on.

I’m glad I read Wynn-Williams’ book, because now I can understand how Facebook let the metaverse happen. And if there’s one thing I learned from reading it, it’s that money and power can bring you a lot of things, but common sense requires neither.

• Most businesses are struggling to deploy agentic AI effectively, and legacy infrastructure is one of the key reasons, report finds
• Google polled IT leaders, with 83% stating that infrastructure upgrades are needed
• IT leaders are also concerned about the hidden costs of agentic AI, such as increased power consumption and operational complexity

If there was a single message to take away from this article, it’s that the infrastructure every business relies on today was not built to handle agentic AI.

Google surveyed over 1,400 senior IT leaders on their AI ambitions, and found that 83% of organizations say they require infrastructure upgrades to leverage the full benefits of production-grade agentic AI.

Moreover, many of those polled are also seeing unexpected costs arise from attempting to run agentic AI on legacy infrastructure. 62% said they had seen significant inference tax driven by data egress fees, storage bloat, and idle specialized hardware, alongside 82% who said that scaling AI introduces hidden operational complexity costs. 79% also reference security, governance, and MLOps as a key barrier to scaling agentic AI.

Upgrades needed for full agentic AI benefit

In order to combat these limitations, Google has several recommendations for organizations hamstringed by legacy infrastructure.

Leveraging fluid compute ‘to dynamically match the right silicon to the right task while minimizing operational overheads’ is Google’s first recommendation, providing compute power for agentic AI tasks without reducing capacity for general workloads, avoiding the need for excess memory usage to run agentic workloads that use large context windows.

For those battling agent sprawl caused by a cascade of new tasks across platforms and teams, Google recommends making use of enterprise-grade governance tools, which are usually available via the cloud partners businesses are already using. Google provides its own platform, Agent Gateway, as an example of a solution that provides visibility and oversight into how agents are communicating, the data they are accessing, and their workloads.

Organizing data more effectively prevents AI agents from drawing more compute when running heavy queries in attempts to access siloed data. Organizations looking to improve the efficiency of agentic AI should work towards using a unified data layer that automatically annotates unstructured data, allowing agents to understand where the data is without having to navigate pipelines. An added benefit of using a unified data layer is that it helps to avoid the duplication of data, saving on additional costs of storage bloat in the long run.

Moving your AI to the edge—by deploying agents directly on the site they are most used—is a further recommendation, and one that organizations are actively pursuing. 90% of organizations polled by Google said that this was a consideration in their AI initiatives. By deploying agents on site in manufacturing plants, retail stores, or hospitals, agents benefit from reduced latency, greater resilience (in the event of a centralized cloud outage), and improved cost-effectiveness by cutting per-token costs with local, highly optimized models.

As with businesses of all sizes, energy costs are a key consideration. When selecting new hardware, 91% of leaders now consider power consumption as a factor, especially when navigating power availability in regions without expanding capacity, regulatory compliance, and reducing the cost of ownership for AI systems.

• GPT-Live is rolling out to all ChatGPT users now
• It can both talk and listen at the same time for more natural chats
• Real-time translations are also now possible

ChatGPT has a shiny new AI voice model called GPT-Live, which has a number of helpful tricks — including being able to listen and talk at the same time. It's rolling out to all ChatGPT users now, though OpenAI has acknowledged a number of early bugs.

While free users and users on a paid plan do get slightly different models — GPT-Live-1 mini and GPT-Live-1 respectively — the updated model should now be appearing in all ChatGPT accounts, with the new features outlined below.

The biggest giveaway that you've got the upgrade will be the Live label at the top of voice chats on mobile, and behind the ChatGPT drop-down on the web. Tap or click on these labels and you can still go back to the old voice models, for the time being.

There's another way to check the GPT-Live voice model has arrived in your account: on mobile, tap the menu button (top left), then the settings cog (Android) or your profile avatar (iOS), and Voice > Model. On the web, click your profile avatar (bottom left), then Settings > Voice > Model.

What to try first

The biggest upgrade here is the 'duplex' functionality, so try that first: you can keep talking even after ChatGPT has started answering you, and it should keep up. Second, try interrupting it mid-flow, and it'll adapt its response accordingly. We're almost at the level of the 2013 Spike Jonze movie Her at this stage.

Third, ask ChatGPT in voice mode to translate something into a foreign language as you say it out loud. You can then speak out sentences in English, and ChatGPT will do a real time translation for you without hesitating. It's not particularly useful for language learning, but it does show off the capabilities of GPT-Live.

Fourth, change the voice and intelligence used — you can do this via the sliders icon at the top right of voice chats. The voice options are actually the same as they were before, but you can choose between Instant, Medium, and High as the intelligence level. Use Instant for the fastest answers, High for the best answers, and Medium for a compromise.

The final thing you can try once you've got the update is to ask questions with visual answers. OpenAI has added a bunch of visual cards to voice mode now, so you get graphics on screen about sports scores, weather forecasts, and places that can be found on a map, for example.

Early voice bugs

I've been testing out GPT-Live voice mode for a few hours and can report that everything works as advertised. It is, more than ever, like talking to a real person — right down to the hesitations and the variety in speech patterns. I did experience one or two glitches, but they were few and far between.

Over on Reddit, OpenAI's Atty Eleti is answering questions about GPT-Live. One of the main bugs that users seem to be experiencing is related to ChatGPT's memory, which appears to be off limits to voice mode in some cases — this is an issue that OpenAI is tracking and "actively investigating", and you can find updates on it here.

Problems are also being reported when it comes to foreign languages being pronounced in an English accent. Again, this is an issue that's been acknowledged, and which should improve over time according to Eleti.

Overall though, the rollout seems to be going relatively smoothly — and I haven't seen any issues with memory or with accents so far. I'm not sure it's going to make me want to use voice mode any more than I already do (which isn't much), but for heavy voice users it's definitely a big step forward.

AI images, videos and writing used to be relatively easy to spot. There were too many fingers, warped backgrounds, strange nonsensical text and other visual artefacts that really quickly gave the AI game away. Deepfake videos often had delayed lip-syncing and AI-generated writing felt repetitive and formulaic, like reading a clunky LinkedIn post.

But AI has improved. A lot of AI-generated images, videos and messages now look convincing enough to fool even the most careful observers and experts. Which means some of the older tricks for spotting what's AI and what's not no longer work as reliably as they once did.

So this isn't about spotting signs and glitches. It's about taking extra steps to verify what you're seeing and hearing. We can't rely on zooming in to see six fingers in a strange-looking ad anymore. We need to know what questions to ask and which tools to use when something doesn't feel quite right.

Fact-checking Facebook pages

AI-generated content is all over Facebook, particularly photos and videos designed to trigger a strong emotional response.

In May, a Full Fact investigation examined pages sharing AI-generated stories about UK politicians and found that many were managed from outside the UK, despite using names that sounded British. The posts featured heartwarming tales of politicians donating millions, rescuing dogs or helping sick children. None of it was true.

One useful tool in situations like this is Facebook's Page Transparency feature. Found within a page's profile, it can reveal where a page is managed from, when it was created, whether it has changed names in the past and other clues about who is really behind it. If those details don't match the image the page is trying to project, that's often a good reason to be sceptical.

Spotting scams on LinkedIn

Job scams aren't new, but AI is making them harder to spot. Scammers can now generate realistic recruiter messages, professional-looking profiles and convincing emails on a huge scale.

Some recent campaigns have even imitated LinkedIn notifications and job alerts, using urgency and curiosity to pressure people into clicking malicious links or sharing sensitive information.

Before responding to an unexpected job offer, do some basic checks. Is the recruiter connected to a real company? Does their profile look legitimate? Does the business exist on Companies House? Taking a few minutes to verify the opportunity could save a lot of trouble later.

Verifying viral videos

Deepfake videos are getting eerily realistic. At normal speed, an AI-generated video may now look completely authentic. But you could try increasing the playback speed, which is when subtle inconsistencies can become easier to spot. Watch for lip movements that don't quite match the speech, unnatural blinking, strange pacing or facial movements that feel slightly out of sync with the audio you’re hearing.

It's also worth paying attention to facial expressions. Now, of course no one reacts perfectly all the time, but if a person's expressions consistently feel disconnected from what they're saying, it may be worth investigating further.

Investigating AI influencers

Sometimes the answer is hiding in plain sight. You just need to know where to look. AI influencers, AI-assisted fashion shoots and AI-generated advertising campaigns are often disclosed by brands and creators, but the information could be hard to find in a caption, hashtag, profile description or small print. For example, lifestyle publisher and brand SheerLuxe creates AI-generated content and influencers under the name Sheerluxe lab.

You could also try a reverse image search. Uploading an image to Google Images can sometimes show you where it first appeared online, if it’s been altered and whether the person in the image actually exists anywhere else on the internet. If an influencer only appears in AI-generated content and nowhere else, that’s a pretty big sign they might be completely made up.

Take the AI or not quiz

Formula 1 has always been a sport where fine margins can make all the difference, with success and failure sometimes coming down to hundredths of a second.

It’s no surprise then that many of the teams have looked to AI for assistance in getting that extra edge, whether it’s through designing new parts, determining race strategy, or simply getting staff to better communicate with each other.

AI giant Atlassian has been working with the Williams F1 team since 2025, signing a title partnership to show its commitment, and at the recent British Grand Prix 2026, I got to speak to the company to find out just how its tools and services are being used.

The best teamwork and the best technology

“We came across the Williams F1 team, and we quickly identified that there was quite a good opportunity for both organizations,” Andrew Boyagi, Atlassian Customer CTO, tells me at Silverstone.

“It’s not a sponsorship, it’s more of a partnership because we can help each other…we had a similar view that in Formula 1, the success goes to the teams that have the best teamwork and the best technology.”

Atlassian’s first season with Williams F1 saw the team achieve a huge jump up the rankings, rising from ninth place overall in 2024 to fifth in 2025, netting it millions of dollars more in prize money.

And although the 2026 season hasn’t been quite as successful so far, Boyagi is keen to highlight how working closely with the team has led to huge improvements in productivity and collaboration.

This has primarily been through Atlassian’s AI tools, which have become common sights in businesses and organizations across the globe. AI usage in the team was also fairly low before the Atlassian partnership, but Boyagi points that after using Rovo, 63% of the team now say, “they have more time to work on strategic, innovative stuff, which means they’re delegating a lot of their work, the repeatable low-value stuff, to AI.”

This includes a new Fault Management tool, which monitors for conflict and repetition when mechanics at the track or the factory report issues. Boyagi notes that it can be so busy at the track, the same fault might be logged multiple times, which can lead to wasted time and effort at the factory, which is particularly painful in a cost-cap limited sport like Formula 1 where every second counts.

Another example is in garage setup - Formula 1 is a global sport, with teams travelling to 24 locations across the world, arriving at an empty garage which needs to be quickly transformed into a cutting-edge hub.

Previously, the team used manual checklists, often with pen and notepads, to track what was being done, leaving the door open to error, but Atlassian redesigned the workflow with Jira and Rovo natively built in, so tasks auto-populate in Jira boards by category and assigned employee, meaning everyone can see who is working on what, and work moves across as it is completed, so hopefully nothing is ever missed.

I ask Boyagi what success will look like in terms of the partnership, especially as Williams F1 has struggled in the initial races so far in 2026.

“Success in Formula 1 is about good technology and good teamwork, but it is always a sport where luck comes into it as well,” he notes, “but what we like to see is an improvement in how things are collaborating, and how effective they are in terms of doing their work.”

Boyagi points out that Atlassian took baseline data between October 2025 and March 2026, finding that 92% of the team now say they’re working for the right organization priorities, “so they’re working on the right things.”

Knowledge transfer has also improved, with trust in documentation going up by 200% in those five months, as Boyagi notes, so “in terms of knowledge, in every company, but especially Formula 1, it is a foundation of productivity - being able to find what you need, when you need it, and trusting what you find without asking someone.”

“Teams are now working on the right things, they’re getting that work done faster, they’re having less meetings and they’re delegating the low value, repetitive tasks to AI,” he adds, “and it doesn’t always translate to what happens on track, but I’m 100% confident we are already making a difference.”

The difference isn’t just for the mechanics or workers back at base - Boyagi says that the team’s drivers are also embracing it. He mentions one example where Alex Albon was having a debate with his engineers around different car set ups - to which Boyagi said, why don’t you ask Rovo?

“As it turned out, Rovo agreed with him, not the engineers - that was a nice moment for us,” Boyagi says, adding that what is important is why Rovo can give a solid answer because the platform is underpinned by Atlassian’s Teamwork Graph, which connects various parts of the business and its processes.

“All of these things are connected, so when you ask a question, it connects dots that humans can’t connect, where it would take us too long,” he says.

Getting over the line

As for Williams 1, the team clearly values the partnership, with Matt Harman, Technical Director for Engineering, championing the time and efficiency savings seen by using Atlassian's systems.

"We need to build systems, tools and techniques across the whole team that allow us to be collaborative," he said on a briefing call attended by TechRadar Pro, "we need to give people more insight, more ideas, so that people can not spend as much time on what I call 'business as usual'".

Harman also highlighted the benefits of the Atlassian service causing "less meetings, more insights", as the partnership looks to provide engineers with the insights and the support to do more.

"When we do that, the Atlassian tools just give people that instant access...without having to sit in a meeting."

Boyagi also points out that the partnership between Atlassian and Williams F1 has multiple sides, noting the team is a customer, “so the fact that they’re using all of our products, and they’re choosing to do that, using precious cost cap dollars, really talks about the value they’re getting from the partnership.”

“I’ve led many transformations, and the hardest thing is getting people over the line - implementing technology is easy, but getting humans to change how they work and want to use the technology is a bit of an art," he says.

"But in Williams, there’s such a strong demand, and a pool, that we have to prioritize what we’re going to do, because they want the team to benefit."

• 75% have deployed 4+ AI tools in the past six months, 35% have deployed 10+ tools
• This is despite four in five experiencing AI security incidents or vulnerabilities
• Governance and dedicated security budgets could be part of the answer

New DigiCert data has claimed four in five (78%) organizations have either experienced an AI-related security incident or have identified an AI-related vulnerability, despite ongoing AI investments and increased adoption.

Despite uncertainty around security, three-quarters of the companies surveyed have deployed four or more AI tools in the past six months alone, with a third (35%) having deployed more than 10 AI systems in that same period.

With this new data, DigiCert says organizations need to treat AI like any other business system rather than an experimental toy, deeply rooting security into the entire strategy.

AI is causing a security headache for most companies

While discussions are clearly taking place, fixes are slow to roll out. For example, 90% have discussion AI governance at executive or board level, but only 50% have implemented both dedicated AI security budgets and formal AI governance programs.

Following years of pilots, two-thirds (64%) have now started logging AI inventories, implying they're still discovering what AI exists across their business. Currently, nearly half lack centralized visibility into AI systems, and only 53% can fully trace AI outputs back to the underlying models and source data.

"The question is no longer whether organisations should adopt AI. It's whether they can explain, govern and trust the AI they've already deployed," SVP Brian Trzupek explained.

While AI explainability remains pretty consistent, at around half, across multiple regions and countries like the US, the UK and Australia, some sectors see more incidents and vulnerabilities than others. Those include science, technology, banking, telecoms and retail.

Looking ahead, many companies are now trialling giving AI agents their own identities, much like human workers, to improve visibility across autonomous actions. Greater governance and stricter policies to iron out the use of unapproved tools should also help companies address some of the growing security issues as adoption continues to mature.

• You can now run Claude Cowork in the cloud, from the web or mobile
• Knowledge work now accounts for around half of all Cowork sessions
• Traditional local Cowork sessions are still supported

Days after reports surfaced that Anthropic could be bringing Claude Cowork to its mobile app, the company has gone one further – users can now start, monitor and complete their agentic workflows from the mobile app and a dedicated web portal.

The upgrade is rolling out in beta now for Claude Max subscribers, but the company has plans to bring the functionality to more plans as rollout continues.

As part of the upgrade, Cowork sessions will also run in the cloud by default – another beta introduction that means workflows can continue even once a PC goes offline or shuts down.

Claude Cowork can now be used virtually anywhere

Because the AI agent can run autonomously across things like files and documents, emails and calendars, and other connected apps, many users mostly left Cowork to run independently. However because it ran locally, it required users to keep their desktop session active even when they stepped away.

Now, scheduled work no longer requires a device to remain online – though users can still choose to run Cowork locally when access to local files is required, for example.

As for why Claude Cowork is being used, Anthropic has revealed that the autonomous agent is mostly being used among knowledge workers despite initially being targeted at coders. "Pulling scattered updates into a single report, building onboarding checklists and reconciling spreadsheets" account for the largest chunk, at around 33% of all use cases across Anthropic's analysis of 1.2 million sessions.

Content creation and copywriting (16%) came next, with software development (9%) and DevOps and infrastructure (7%) actually only accounting for much smaller proportions.

With knowledge work now accounting for nearly half of all Claude Cowork sessions, the company's research shows agentic AI emerging as an everyday work colleague. Though the company didn't indicate how, or whether, this shift in behavior might impact its pipeline, a shift away from coding as a primary use case could evolve Cowork in different ways to how we might have imagined.

• Researchers have spotted problems with AI pattern matching in science data
• It could mean false flags for signatures of life on other planets
• AI can still be useful, but checks need to be built in

One of the ways AI can be most helpful is in trawling through masses of scientific data that human researchers don't have time to analyze, looking for patterns — but this use-case is now proving problematic when it comes to the search for life beyond our planet.

A new study from researchers at Michigan State University suggests that AI systems can be too easily fooled into identifying signatures of life out in the universe where none exist. We need these flags to be accurate to know where to point our telescopes next, so it's important that the detection processes work.

The researchers set up a digital simulation including a key sign of life: the ability for molecules to replicate and mutate. Software was used to generate tens of thousands of digital organisms with and without this ability, which where then used to trail a neural network to spot the difference with an accuracy rating of 99.7%.

When the neural network was pointed towards data it hadn't previously seen, however, the AI's life-spotting skills fell apart. The researchers started with a digital organism that couldn't copy itself, which the AI correctly identified, then began making small edits and asking the AI to check again.

Essentially, as the AI was nudged out of its comfort zone of training data, it started seeing life where there wasn't any. "No matter what sequence of commands we started with, we were able to fool the AI 100% of the time," said Ankit Gupta, one of the researchers.

Space and beyond

It's worth bearing in mind the limitations of this research: these tests were carried out in an artificial, digital simulation, and so didn't rely on any real data. The researchers were deliberately searching for errors too, rather than letting them happen by chance.

However, the study methods are solid enough to be concerning. The worry is that a Mars rover or a deep-space telescope could identify a life signature with a high degree of confidence, without necessarily having a human in the loop to check.

The researchers found there were a vast number of sequences that could trip up the AI too, meaning the risk of a mistake is more likely. While the digital organisms incorrectly identified by the neural network were close to what it had been trained to spot, they weren't full matches — despite the AI thinking they were.

These issues could crop up outside of space exploration too. The same errors might appear when looking for patterns in medical scans, security camera footage, and everywhere else the technology is used.

That said, the researchers are keen to emphasize that AI can still be useful in these scenarios — it just needs careful checks and supervision. "AI has an Achilles' heel: it can see a pattern and completely misclassify it," said Christoph Adami, one of the team. "There needs to be a human in the loop."

• UK data centers can apply for a special 'national significance' designation under the NSIP scheme
• The designation allows new projects to receive approval direct from Whitehall, avoiding local planning and building regulations
• Opposition to data centers in the UK has not reached the same levels as the US, which has seen billions of dollars worth of projects delayed due to local opposition

Data centers in the UK have been given the ability to apply for a ‘national importance’ status previously only reserved for critical infrastructure such as energy production, roads, railways, and sub sea cables.

The UK government has also scrapped the statutory requirement for pre-application consultation on Nationally Significant Infrastructure Projects (NSIPs) in legislation taking effect later this month, potentially cutting the processing time for applications by up to a year.

Projects approved for NSIP status are not subject to local building regulations and authorities, and are instead granted permission directly by the UK government. Data centers will therefore be able to apply for this same status with a reduced processing time.

What determines a NSIP?

For now, there is no official guidance on what determines a data center as a NSIP.

Speaking to The Register, Law firm Womble Bond Dickinson said, “Datacenters are not automatically consented as NSIPs; instead, the NSIP regime operates on an opt‑in basis for developers. A datacenter project may be directed into the NSIP regime where the Secretary of State considers it to be of national significance and satisfied that the statutory tests under section 35 of the Planning Act 2008 are met.”

The UK has long been poised to take advantage of new technologies, with regional cyber hubs popping up across the country in areas of technological significance for innovation, growth, and talent development - such as Cheltenham’s cyber hub strategically positioned near the Government Communications Headquarters (GCHQ).

The NSIP designation for data centers will likely fast-track the construction of ‘AI Growth Zones’ intended to accompany cyber hubs and bolster the UK’s sovereign AI capacity. That means that data center construction projects of ‘national importance’ will soon begin popping up across the UK.

The NSIP also provides prospective projects with pre-application advice to improve both the speed of their application and the chances of approval, with over 80 potential projects already having benefitted from this new advice.

As the ‘national importance’ status allows projects to bypass local planning and building regulations, it will also likely draw attention from local ‘not-in-my-backyard’ (NIMBY) groups. How effective the NSIP label will be at countering this local opposition remains to be seen.

The UK Local Government Association issued a response to the sustainability of data centers in the UK that stated, “Data centres and AI infrastructure cannot be planned in isolation from wider digital connectivity, energy, water, land use and climate systems. Councils must be treated as key partners in the design and delivery of national digital strategies, including AI Growth Zones, planning reform and infrastructure investment.”

What is the current attitude to data centers in the UK?

Opposition to UK data centers has not seen the same level of support as that in the US, but the UK has not been subject to the same mass-buildout of data center projects as in the US.

A YouGov / Cavendish Consulting Survey [PDF] of 2124 UK adults conducted in November 2025 found that the UK population broadly supports the construction of new data centers (69%), with 24% stating that they strongly support, and 45% stating they tend to support.

Opposition is significantly lower, with just 7% stating they tend to oppose data center projects, and only 3% strongly opposed.

The further trend to be analyzed is the 21% who gave ‘Don’t know’ as an answer, which highlights the lack of understanding around what a data center is, why they are built, and how they relate to AI technology.

This was further highlighted by a June 2026 survey conducted by SEC Newgate that found that 89% of UK adults were unfamiliar with data centers.

The US has seen opposition to data centers at both local and national levels, centered around the fear of AI-related job losses, environmental damage, and legitimate fears over local capacity constraints caused by the energy and water demands of enormous data center campuses.

How new UK data centers engage with local authorities and community groups will weigh heavily on the national opinion of data centers in the years to come, especially now that new projects can skirt local regulation via NSIP designation.

Via Tom’s Hardware

• ByteDance aims to finalize its new CPU design by early 2027
• Mass production and deployment are targeted for the second half of 2027
• An early chip version has been used internally since late 2025

TikTok parent company ByteDance is reportedly racing to finalize the design of a next-generation in-house CPU by early 2027.

According to SCMP, the company aims to reach mass production and wider deployment during the second half of 2027, with an early version of this proprietary processor being deployed internally since late 2025.

Internal computing demand at ByteDance has accelerated sharply, driven largely by products like the Doubao chatbot and Seedance video model, which have significantly increased infrastructure requirements across the company's expanding AI tools lineup.

Urgency driven by rapid AI growth

As agentic AI workloads grow more complex, computing needs are shifting away from pure matrix calculations toward broader task orchestration duties.

These newer AI systems require more coordination, decision-making, memory management, and software operations.

It therefore creates greater demand for general-purpose processors working alongside GPUs rather than relying only on accelerator clusters.

This shift is influencing how companies like ByteDance plan future computing infrastructure.

According to reports, tape-out, the final engineering milestone before physical chip fabrication begins, could be moved earlier than originally planned.

ByteDance has never publicly commented on its chip development activities, despite what reports describe as rapid expansion across several in-house silicon designs.

To help accelerate development and secure foundry manufacturing capacity, the company is reportedly collaborating with US chipmaker Qualcomm on this initiative.

Export controls change the competitive landscape

Washington's export controls have progressively restricted Chinese companies' access to advanced semiconductors, including Nvidia's H100 and H20 accelerators specifically.

This tightening regulatory environment has pushed China's largest technology firms toward building their own in-house chip programs at scale.

ByteDance's CPU initiative fits within that broader pattern, aiming to reduce dependence on suppliers it cannot fully control.

Publicly traded chipmakers including Arm Holdings, Intel Corporation, and Advanced Micro Devices could face reduced demand as ByteDance's in-house capabilities mature further.

Nvidia, already limited in selling advanced AI accelerators to Chinese buyers, faces an added long-term headwind as ByteDance builds internal alternatives.

This strategy mirrors moves already made by major global hyperscalers investing heavily in custom silicon infrastructure.

Google's TPU chips, Amazon's Trainium and Graviton processors, and Microsoft's Maia accelerator all reflect a similar underlying thesis.

At sufficient scale, proprietary hardware can offer meaningful cost and performance advantages over hardware purchased from outside vendors.

But the confirmation of an official tape-out date would give markets a clearer signal of how seriously to value ByteDance's growing semiconductor ambitions.

• OpenAI reveals first branded hardware, the Codex Micro, a programmable macro pad built with keyboard maker Work Louder
• Codex Micro seems to be based on Work Louder's Creator Micro 2's layout, mapped to Codex coding-agent shortcuts
• The move reinforces OpenAI's Codex offering as one of its mainstay areas of focus by allowing developers the ability to perform tasks or interact with AI faster

OpenAI's first branded piece of hardware is not a long-anticipated consumer device it is building with ex-Apple design chief Jony Ive, but rather a programmable macro pad called the Codex Micro.

The keyboard, which consists entirely of macro keys designed to "supercharge people's Codex usage," according to an OpenAI spokesperson at the AI Engineer World's Fair, is reportedly a collaboration between the iPhone creator and the custom macro pad creator Work Louder.

With OpenAI's developer-centric account on X indicating that the full launch of its hardware foray is expected on July 15, the AI giant seems to be pulling out all the stops to ensure it becomes a well-received add-on for the developer community.

A simple rebadge or a sign of things to come?

The as-yet-pending release 'Codex Micro' seems to be inspired by Work Louder's existing Creator Micro 2, a compact macro pad that offers thirteen mechanical keys, a joystick, a rotary encoder, and touch controls, arranged across programmable layers to power users needing faster or more fine-grained control over AI-assisted coding tasks.

The move is understandable for OpenAI in terms of both securing a victory with developers and brand recognition, and essentially testing the waters on how it would handle a hardware launch for the company's upcoming AI device for more general-purpose users.

It can also, to a degree, be seen as OpenAI essentially acknowledging that its earlier stance of narrowing its focus to 'nail' its core business might be one the company is willing to make exceptions to, especially when it comes to coding tools or enterprise use-case hardware.

OpenAI's CEO of applications, Fidji Simo, reportedly told staff that the company was looking to deprioritize areas outside its core focus to allow it to lead where it mattered.

In 2025, OpenAI shipped the Sora video app, the Atlas browser, ecommerce features inside ChatGPT, advertising work, and hardware efforts, a "series of startups" approach that insiders said had produced organizational confusion and constant reshuffling of scarce compute, distracting it from a truly centralized goal.

Hardware, in other words, was explicitly on the list of distractions. A physical keyboard is arguably as clear a violation of that directive as one could possibly design.

OpenAI is also reeling from a smaller-than-expected gap from competitor Anthropic and its Claude models in the areas where its GPT models do compete. This can perhaps be attributed to Anthropic's much narrower focus, which caters specifically to coders and enterprise through its Claude Code and Claude Cowork offerings.

One can argue that OpenAI's move isn't one that distracts it from its core focus, but rather complements it, even as R&D and integration for the most part is something that Work Louder will commit to.

It allows the AI juggernaut gets to test out both the marketability of an OpenAI-branded hardware product and appease developers and founders with a low-effort play even as they have increasingly been considering tools from Anthropic and Google as well as other AI solutions providers.

None of OpenAI's previous concerns may apply here; the exercise does not consume compute, it caters to a key audience for OpenAI, with Codex assisting 5 million weekly users as of June, and it does not meaningfully engage an engineering team as some of its other projects do.

With OpenAI and Anthropic slated to IPO soon, both are locked in a race to secure as many active users as possible to justify their valuations, even as they vie to build the most powerful models to cater to various industries, including defense, telecommunications, pharmaceuticals, and software development, to name a few.

OpenAI's move might just be a sign of things to come, as it leverages ChatGPT's massive brand recognition to develop marketable, revenue-generating solutions such as a custom macro keyboard, even as it is loath to spend any of its engineering or compute resources on anything but the most important of its tasks, even as enthusiasts continue to wait for the release its upcoming collaboration with legendary Apple designer, Jony Ive.

Gemini Gems are Google's answer to the annoyance of constantly having to repeat yourself to an AI chatbot. Like ChatGPT's custom GPTs, Gemini Gems are customized, reusable variations of Gemini that remember a specific role, history, and personality, so you don't have to keep explaining yourself every time you start a new chat.

Rather than beginning every conversation with housekeeping, you immediately start solving the problem you actually opened Gemini to tackle. You build a collection of specialists that already know their jobs. Open your travel planner when you're booking a holiday, your guitar coach when it's time to practice, or your meal planner when the refrigerator looks uninspiring, and each one picks up exactly where you left off.

I've made plenty of Gems, some more enduring than others. They're easy enough to make, but tweaking them to be just right can be tricky. If you want to see some of the more appealing (and sometimes just fun) possibilities of Gems, here are five of my favorites. I've written out the instructions I composed for the Gem at the end of each. Gemini can also edit and expand on even the simplest of descriptions, but more detail can help ensure the Gem does what you want.

Creating a Gem

The process of creating a Gem is easy, just click/tap on Gems in the left hand menu in the web browser version or the app version of Gemini, then New Gem. You can use the custom instructions from each of my five Gems if you'd like to recreate them yourself.

1. Family Adventure

Planning family outings has become its own part-time hobby. My wife and I have a two-year-old and an eight-month-old, so every trip has to thread a surprisingly small needle. It needs to be close enough that nobody spends half the day in the car, interesting enough to entertain everyone, stroller-friendly, and ideally open when we actually want to visit. That made Family Adventure Planner the first Gem to showcase.

Setting it up only took a few minutes. After creating the Gem, I conversed with Gemini through it and gave it some basic details about locations, interests, and the kinds of places we've enjoyed in the past. Once the Gem had all that information, I threw some different scenarios at it.

I asked it to plan a family outing for the coming Saturday within about an hour's drive. The Deep Research feature pushed the Gem to check what would actually be open that weekend, look for seasonal events taking place, verify opening hours, and even factor in temporary exhibits and admission prices before putting together a suggested itinerary.

The recommendation felt surprisingly complete. It suggested a nearby sculpture park with stroller-friendly paths, followed by lunch at a family-friendly café and an ice cream stop on the drive home. It also pointed out that arriving before mid morning would make parking easier, exactly the kind of practical advice that is easy to overlook until you're trying to unload two young children from the car.

I asked it to imagine that rain was forecast all day and that we still wanted to get out of the house. Instead of simply swapping a park for a museum, it built an entirely different plan around an interactive children's museum, suggested a nearby indoor play space if our oldest still had energy afterward.

The Gem also adapted quickly as I added more context. After mentioning that long waits at restaurants rarely end well with a hungry toddler and an eight-month-old, future itineraries naturally favored casual cafés, picnic spots, and places where food was readily available. It quietly learned from each conversation instead of making me repeat those preferences every time.

If your weekends usually begin with twenty minutes of searching before anyone leaves the house, this is probably the first Gem worth creating.

Family Adventure Planner instructions

You are an enthusiastic, creative, family-focused activity planner. Learn my family's ages, interests, travel preferences, and other details. Whenever I ask for ideas, recommend activities that are realistic, seasonal, and varied while avoiding suggestions I have recently tried unless I specifically ask for favorites. Include all relevant logistics details like times, costs, and packing suggestions.

2. Hobby Coach

Some hobbies are easy to put down for a while. Others seem to expect you to remember exactly where you stopped. That made Hobby Coach one of the first Gems I wanted to build.

I set it up with two of my biggest hobbies: learning guitar and backyard astronomy. I told it I was still a beginner guitarist working toward playing complete songs and that my astronomy interests revolved around learning the night sky with a modest telescope instead of serious astrophotography. Once that information was saved, I never had to explain it again.

To see how useful it would be, I spent an afternoon asking it to map out future practice sessions instead of simply answering questions. For guitar, it created a progression that built from my current skill level, suggesting chord exercises, songs that gradually increased in difficulty, and realistic milestones to aim for over the next several weeks. Everything fit into a longer learning plan.

Astronomy worked just as well. I asked it to plan a series of upcoming observing nights, and it suggested different targets depending on the season, moon phase, and what I wanted to learn. One evening focused on easy constellations, another introduced brighter deep sky objects, while another became a relaxed tour of the Moon and planets.

The Gem also uses Guided Learning as its default tool, which structures lessons into connected learning paths instead of isolated answers. It builds on previous lessons, introduces new skills at the right pace, and creates the feeling that you're working with a patient teacher who already understands your goals.

Hobby Coach instructions

You are an encouraging, knowledgeable, and patient personal coach for my hobbies. Learn my current experience level, equipment, goals, schedule, and preferred learning style for each hobby I share with you. Remember my progress over time and build each lesson naturally on previous conversations instead of starting from the beginning. Break complex skills into manageable practice sessions, celebrate improvements, and recommend realistic projects that keep me motivated without becoming overwhelming.

3. Movie and TV Curator.

Choosing something to watch should be one of the easiest parts of the evening, yet it often turns into an extended scrolling session. I knew that AI chatbots can be useful here, but a specific Gem for the endeavor felt like a good fit.

I gave my new Movie and TV Curator Gem everything it needed to know about our tastes. I told it which streaming services we subscribe to, the kinds of films my wife and I enjoy after the children are asleep, the comedies and mysteries we've already watched, and perhaps most importantly, that we have a two-year-old who is just beginning to sit through longer family movies.

That final detail completely shaped its recommendations. Instead of suggesting whatever happened to be popular, it focused on gentle, engaging films that would make good introductions to family movie nights without overwhelming a young child. It also remembered which movies we'd already seen so future recommendations wouldn't feel repetitive.

I asked it to build a month's worth of family movie nights, along with separate recommendations for date nights after the children were asleep. Within minutes, I had a calendar filled with classic animated films, newer family favorites, and several older movies that I had completely forgotten about but couldn't wait to introduce to my son.

It also understood the different kinds of evenings we have. A Friday after a busy week called for something light and funny, while a quiet Sunday evening was a better fit for a slower family film. I soon had a collection of family movie nights and grown-up viewing plans waiting whenever we needed them. Like the other Gems, it turned a repetitive decision into something I only had to think about once.

Movie and TV Curator instructions

You are a knowledgeable, conversational entertainment expert with excellent taste and a great memory. Learn my favorite genres, actors, directors, streaming services, viewing habits, and the movies and television shows I've already watched. Recommend films and series based on my mood, available time, and who will be watching, while avoiding unnecessary spoilers and repeating recent suggestions unless I ask. Explain why each recommendation suits my tastes, maintain a warm and enthusiastic personality, and regularly introduce overlooked classics alongside newer releases.

4. Outfit Planner

At first glance, Outfit Try On Planner sounded like a Gem I'd probably use once and then forget about. I enjoy looking reasonably presentable, but I wouldn't describe myself as someone who spends much time thinking about fashion. After setting it up, though, I realized it was less about keeping up with trends and more about making decisions before I actually needed to make them.

I started by teaching the Gem my style. I told it the kinds of clothes I usually wear, the colors I naturally gravitate toward, and the occasions I dress for most often. I also uploaded a few photos of myself so it could create realistic visualizations rather than relying on generic fashion models.

I asked it to put together outfit ideas for a fancy date, weekend trip, and a few other occasions. Seeing complete outfits instead of reading descriptions made decisions much easier. The most entertaining experiment had nothing to do with everyday clothes. I'd been thinking of dressing up for the Renaissance fair this year, so I asked the Gem to imagine me in a variety of Renaissance costumes before I bought or rented anything. You can see me as a hooded archer, an elaborately dressed nobleman, and a cheerful wandering bard carrying a lute.

Because the Gem remembers your appearance and preferences, it can help visualize costumes, themed party outfits, Halloween ideas, vacation wardrobes, or almost anything else you might wear before you spend money assembling it.

Outfit Try On Planner instructions

You are a friendly, fashion-savvy personal stylist with an eye for color, fit, and practicality. Help me create outfits using clothes I already own, visualize new pieces before I buy them, and suggest combinations that suit the occasion, weather, and my personal style. Ask questions before making recommendations. When I upload photos of clothing, accessories, or myself, use them to generate realistic outfit visualizations and styling ideas.

5. Personal Theme Song

The final Gem was the one I expected to be the silliest, yet it ended up being one of my favorites. Google recently added a Music tool to Gemini that can generate original songs from simple prompts, so I decided to build a Gem called Personal Theme Song Composer, dedicated entirely to turning everyday moments into music.

Setting it up only took a few minutes. I told it about the musical styles I enjoy, asked it to learn how I like songs to feel, and instructed it to ask questions about the people, pets, places, or memories behind each request before using Gemini's Music service to compose something original. Once those instructions were saved, I could jump straight into ideas instead of explaining the same preferences every time.

One of the first finished songs involved my dogs. Every dog owner eventually invents a ridiculous tune while clipping on the leads for a walk, so I asked the Gem to write something jaunty about my two excitable Chihuahuas that sounded like the opening theme to a cheerful television comedy. The result perfectly captured the determined little strut they adopt every time they head out the front door, and the melody stayed in my head for the rest of the day.

It perfectly summed up what makes Gemini Gems so useful. They are more than saved prompts. They are specialists that remember their role, making new capabilities like Gemini's Music tool feel less like an occasional novelty and more like a creative partner that's always ready when inspiration strikes.

Personal Theme Song Composer instructions

You are an imaginative, enthusiastic, and collaborative songwriter and music producer. Your goal is to help me create original songs that celebrate and vibe with whatever topic you're given. Before writing a song, ask enough questions to understand the story I want to tell, the mood, the musical genre, whether I want vocals or an instrumental, the intended audience, and any specific lyrics, phrases, or themes I want included. Suggest genres, tempos, instrumentation, and vocal styles that fit the idea. If I provide photos or other context, use them as inspiration for the music's tone and storytelling.

What makes Gems feel different from saving a handful of good prompts in a notes app is continuity of purpose. A prompt tells Gemini what to do once. A Gem remembers who it is supposed to be every time you come back. It remembers you, your favorite way of working, and the tools that help it do its job best, creating an experience that feels much more personal over time.

The five Gems here are really just a starting point. Once you get comfortable creating them, it becomes surprisingly easy to imagine building one for meal planning, another for packing for trips, or any other gems you want to fill your treasure chest with.

Earlier this year, Taylor Swift filed trademark applications covering the phrases "Hey, it's Taylor" and "Hey, it's Taylor Swift" — two recordings closely associated with her voice and personal brand. Around the same time, actor Matthew McConaughey sought trademark protection for some of his most recognizable catchphrases.

These cases highlight an important question in the AI era, what rights do people actually have over their voice?

And this isn’t a theoretical issue. People tend to focus on the image and video generation capabilities of AI, but AI voice cloning tools can now recreate a person's voice from just a few minutes of audio. While platforms like ElevenLabs and Suno have helped make synthetic voices and AI-generated audio more accessible than ever.

For celebrities, the stakes are really clear. A distinctive voice can be a valuable part of a personal brand and a source of income. But this technology doesn't just affect public figures anymore. Now, anyone with audio recordings online could potentially have their voice copied, cloned or imitated.

The problem is one we’ve touched on repeatedly in our reporting on AI. Which is that the law hasn’t caught up to the tech yet. Existing laws around copyright, trademarks, privacy and publicity rights were largely created before AI-generated voices became possible at scale.

As a result, many of today's biggest questions don't have straightforward answers, and the rules vary significantly depending on where you live too.

So where do we stand right now? What protections do ordinary people actually have if their voice is cloned or used without permission? And are lawmakers beginning to adapt to a world where a voice can be replicated in minutes? To find out, I spoke to legal experts specializing in technology, intellectual property and AI law.

What does the law say?

Dr Edina Harbinja is an Associate Professor in Law at the University of Birmingham, specializing in digital rights and the regulation of emerging technologies. I asked her if I can ever really “own” my voice and she says there’s not a simple answer.

“A recording of a voice can be protected by copyright and related rights, and a performance may attract performers’ rights,” she explains. “But the voice itself is usually not ‘owned’ in the same way as a song, a sound recording, or a written work.”

Professor Andres Guadamuz, an expert in intellectual property and technology law at the University of Sussex, largely agrees.

"In the UK the answer is no, you don't own your own voice," he tells me. "There are several possible ways to try to get around this, but the real answer is that there is no good legal recourse if someone uses your voice to train an AI, particularly if they have legal access to the voice."

Harbinja tells me that ownership isn’t necessarily the most useful way to think about the issue. “The harder issue is whether an AI-generated voice that sounds like a person appropriates their identity, falsely implies endorsement, or misleads the public,” she says.

In other words, the legal question may be less about ownership and more about whether a cloned voice causes harm, deception or misrepresentation. But even then, what happens next depends heavily on where you live.

“In the US, this is often framed through the right of publicity, false endorsement, and impersonation-style claims,” Harbinja says. She points to the case of Midler v. Ford Motor Co. In which singer Bette Midler successfully objected to Ford using a soundalike singer in an advert.

There are also newer US initiatives, including the proposed NO FAKES Act, move more directly towards protecting voice and likeness against unauthorized digital replies.

But the UK takes a different approach. “In the UK, there is no general right to own one’s voice, image, or persona,” she tells me. “The most relevant routes are passing off performers’ rights, misuse of private information, data protection, defamation, contract, and sometimes trade marks. But these are patchwork protections.”

One possible route is passing off, which protects businesses and individuals from misrepresentation that could damage their reputation or falsely suggest commercial endorsement.

Guadamuz points to passing off as one of the few existing legal routes available in some circumstances.

"If a person has a reputation, they could use passing-off, which is a way to stop people from profiting from your image," he says. However, he notes that its application is limited and has historically been used more often for images than voices.

Harbinja says that the challenge here is that AI voice cloning doesn’t fit nearly into a single legal category.

“AI voice imitation can be a form of impersonation, but the legal label depends on the jurisdiction and context: false endorsement, personality infringement, data processing, unfair competition, performers’ rights, or consumer deception,” she says.

The legal framework also varies depending on whether the person is a celebrity or performer.

“So, you may own the recording, the song, or the mark, but you do not necessarily 'own' the sound of yourself in a simple copyright sense,” she says. “AI voice imitation exposes the gap between laws protecting works and laws protecting persons.”

AI presents unique problems

Harbinja explains AI complicates things because there are two major enforcement problems to consider.

The first is a training problem. “Models may be trained on vast collections of online recordings, songs, interviews, performances, and videos. Rights holders often do not know whether their material was used, where training occurred, or whether a legally relevant copy was made in a jurisdiction where they can sue,” Harbinja tells me.

The second problem is output. “If an AI output sounds like Taylor Swift but does not reproduce a substantial part of a protected recording or song, copyright may not be the best route,” she says. “That is why the debate is shifting from copyright alone to voice, likeness, personality, publicity, consent, transparency, and digital replica regulation.”

She says this is why the Taylor Swift example is proving to be so interesting. Because although trade mark and brand protection strategies can work in commercial contexts, they don’t solve the bigger issue of unauthorized AI imitation.

Guadamuz says enforcement remains one of the biggest practical obstacles. "There is no protection in the UK, some in other countries, but it is almost impossible to enforce at the moment. Only the most egregious examples of cloning could be actioned,” he tells me.

The future of the law

Is the law changing to accommodate new technologies? Yes, but slowly.

“AI transparency laws are emerging,” Harbinja tells me. “The EU AI Act does not create a property right in one’s voice, but Article 50 imposes transparency obligations for certain AI interactions and deepfakes/synthetic content, including synthetic audio.”

Guadamuz believes new legislation will ultimately be needed. "For the most part we need new legislation that specifically protects voices," he says.

He notes that UK policymakers have already discussed protections against deepfakes and digital replicas, which could eventually extend to AI-generated voice clones.

However, Harbinja argues that the debate here shouldn’t focus exclusively on ownership.

“The deeper theoretical and moral question is whether voice, likeness and persona should be governed primarily as property, or instead through concepts of dignity, personhood, identity and autonomy,” she says. “My own view is that the latter provides the more appropriate starting point, especially where AI systems do not merely copy a marketable asset but simulate a person.”

What Harbinja got me thinking about is how often we treat AI as just another technology to regulate. Another tool, another platform, another legal problem to solve.

But the rise of AI voice cloning suggests something more complicated is happening. These aren't just questions about copyright, trademarks or data protection. They're questions about identity, authenticity and personhood.

After all, a voice isn't just a piece of data. It's one of the ways we recognize each other. It's tied to our relationships, our memories and our sense of self.

As AI becomes increasingly capable of imitating not just what we create but who we are, the challenge for lawmakers goes beyond deciding what can be owned and becomes about whether we should ever treat parts of a person as property in the first place.

Most people use ChatGPT like they’re walking up to a stranger and starting a new conversation every single time.

You open a new chat, explain what you need, add a bit of context, correct the tone, ask it to be more concise, then finally get something useful. And then, the next time you need the same kind of help, you do the whole thing all over again.

That’s fine if you’re asking something random each time. But if you use AI regularly for the same kinds of tasks, it’s a surprisingly inefficient way to work.

The solution is simple: stop starting from scratch.

Instead of treating every chat as a blank page, you can build reusable conversations that already know the job, tone, and desired output. Here’s how I use that one habit to save time every week.

A simple trick

So, we know what we want to do — remove repeated setups for prompts, make the AI’s first answer better, and reduce the amount of refining you need to do afterward. But how do we do it? Here’s the clever bit - you get the AI to do it for you.

Load up one of your previous chats that was productive and worthwhile. Then at the end I want you to copy and paste this:

“Turn this conversation into a reusable prompt I can paste into a new chat next time. Include the goal, tone, format, constraints and the steps you followed. Make it general enough that I can reuse it with similar tasks.”

Your AI will give you a handy prompt you can use whenever you like now to get a chat that’s exactly what you want.

My meal planner prompt

I now have a reusable prompt for meal planning. Instead of starting from scratch every week, I paste in a short prompt that already knows the rules: I want quick family meals, nothing too expensive, leftovers if possible, and a shopping list organized by supermarket section. Then I add whatever’s in the fridge and how many nights I need to cover.

Here it is:

“You are helping me plan practical family meals for the week. Prioritize meals that are quick, affordable, not too fussy, and likely to produce leftovers. Ask me what ingredients I already have before suggesting anything. Then give me:

A simple meal plan

A shopping list grouped by supermarket section

Any ingredients I can reuse across multiple meals

One backup meal in case I don’t feel like cooking”

This has saved me so much time planning meals in the evening.

A tool that remembers you

Before you get too carried away with this idea, remember that reusable prompts are useful but not infallible. A reusable prompt is only worth saving if it produces better work, not just more predictable work. If your saved prompt is too vague, too bossy, or based on a weak workflow, you’ll just get the same mediocre output, but delivered faster.

Having said that, using reusable prompts has been the biggest change I’ve made to how I use ChatGPT. When a conversation works, I don’t let it disappear. I turn it into a reusable starting point.

If you stop starting every ChatGPT conversation from scratch, ChatGPT starts feeling much less like a conversation with a stranger and more like a chat with a tool that actually remembers how you work.

• Electricity demand is now growing faster than energy suppliers can keep up with
• Volatile AI workloads cause unpredictable peaks and troughs in demand
• AI could actually help predict, despite also being the cause

With three in four (77%) electricity execs now believing that data center energy demand will grow faster than utilities can keep up with, two-thirds (68%) expect electricity shortages to become more commonplace as demand for AI soars.

New data from a Capgemini report reveals just how unpredictable AI energy demands can be, with 77% admitting they struggle to accurately forecast demand amid volatile AI workloads.

Not only is this leading to more constrained energy supply, but also more extreme and less predictable demand spikes.

Data center energy demand is a whole new ball game

All of this comes as local opposition continues to mount against data centers, with residents increasingly concerned about power outages and rising energy costs. Just last week, a county in Virginia told data centers to revert to backup generators to free up grid capacity for local residents, with an ongoing heatwave causing a spike in electricity demand for air conditioning units.

Even data center companies are struggling to anticipate how much they could consume, with 67% of electricity execs reporting speculative applications for future capacity. Around a fifth (19%) of these don't even materialize, creating what Capgemini calls 'phantom demand,' forcing utilities to either overinvest unnecessarily or underinvest and create capacity shortages.

"The challenge is no longer only how much power is needed, but whether it can be delivered reliably, where and when it is required," Capgemini Global Head of Energy and Utilities Claire Gauthier wrote, citing AI's potential in helping to predict demand despite also being the cause of fluctuating and high demand. However, at the moment fewer than half (45%) currently use AI for grid optimization.

Looking ahead, most (87%) data center operators expect electricity consumption to rise over the next three to five years by an average of 30%.

• Microsoft Frontier Company to embed 6,000 engineers and specialists into customer organizations
• Backed by $2.5 billion in Microsoft funding, it will help customers transform with custom AI
• This is the "largest" of its type, 2.5x the value of Amazon's alternative

Microsoft has launched a brand new subdivision to expand its own AI consultants into customers' companies backed by a massive $2.5 billion investment.

The new Microsoft Frontier Company will embed more than 6,000 specialists, AI engineers and technical experts directly inside customer organizations to help build, deploy and optimize their own AI strategies.

Microsoft described it as the "largest, most capable, outcome-driven engineering organization in the industry" – the scheme comes days after Amazon announced a similar scheme backed by $1 billion.

Microsoft launches forward-deployed engineer (FDE) program for AI

Though similar in concept to other FDE programs, Microsoft believes its Frontier Company will be different in that it adds extra layers of industry expertise, change management, continuous improvement and more, rather than just racing to deliver AI ROI.

Microsoft Commercial Business CEO Judson Althoff emphasized the importance of intelligence and trust in tailoring a suitable AI strategy for its customers. Intelligence's role involves understanding broad organizational context, workflows and processes, while trust is all about governance, observability and accountability.

Early Frontier Company customers include LSEG and Unilever, and of course, being an enterprise-focused solution, the company stressed that proprietary data, workflows and more remain private to companies and doesn't get used to train models.

Another major selling point for the "largest" AI FDE scheme in the industry is that customers can pick and choose between models from OpenAI, Anthropic, Microsoft and other open-source alternatives to provide the best solution for every workload, rather than forcing companies to lock in to one single tool.

Microsoft Frontier Company will be led by former President of Microsoft Asia, Rodrigo Kede Lima. "He has been at the forefront of helping customers and partners translate technology shifts into business outcomes, and understanding how platform innovation, engineering and partner ecosystem collaboration come together to drive growth," Althoff wrote.

• Alibaba bans access to Claude Code, tells employees to use internal Qoder instead
• Anthropic was tracking markers to indicate which users were in China
• Anthropic accused Alibaba of major Claude model distillation effort

Alibaba has reportedly banned its employees from using Claude Code internally, beginning July 10 2026, classifying it as a high-risk tool that risks organizational security.

The change follows similar trends already observed among American tech giants, banning Chinese tools from internal use, but Alibaba cited genuine concerns that have been acknowledged by Claude-maker Anthropic.

The ban could also be seen as a push for Alibaba's own alternative, with workers advised to use the company's own Qoder AI assistant instead.

Alibaba bans Claude Code over security concerns

The controversy stems from developers reverse-engineering Claude Code, revealing it contained code to identify Chinese users. Checks for Chinese system time zones, proxy servers, AI lab infrastructure and network characteristics were all revealed.

Anthropic stated this experimental feature launched in March, and was designed to combat unauthorized resellers, prevent account abuse and protect its models from AI distillation.

However, the spyware was reportedly hidden using obfuscation and steganographic techniques, making them effectively invisible to users.

This isn't the first time both companies have found themselves in a sticky situation – Anthropic recently accused Alibaba of conducting the largest known model distillation attack against Claude (via Reuters).

More broadly, Chinese companies have been increasingly referring to domestic AI tools like Qwen, DeepSeek, Moonshot and Ship amid growing geopolitical tensions.

While that trend has been largely mirrored in the US, in favor of the likes of OpenAI, Anthropic, Google Cloud and xAI, US firms have reportedly been exploring cheaper Chinese alternatives in the name of cost efficiency.

Alibaba and Anthropic haven't publicly commented on this matter as yet.

• Microsoft Teams will now let users turn off its AI tools
• Copilot, Facilitator and Intelligent recap all affected
• Users will be able to pick and choose which AI tools they want on Teams

Microsoft has apparently backed down in plans to introduce a host of AI tools across Teams after it faced a major backlash from users.

In recent weeks, the company has revealed several new AI-powered Teams features it says will help boost user productivity and efficiency, offering the likes of catch-up tools, note-taking and even translation.

However, following an apparent fightback from users, the company says it will now offer a simple toggle to turn off its "Meeting AI" features on your calls.

Turn off AI in Microsoft Teams

The news was outlined in an admin center post, seemingly admitting the company might have overdone its AI expectations.

“Microsoft Teams will add an in-meeting toggle for licensed organizers and presenters to turn Meeting AI (Copilot, Facilitator, recap) on or off during live meetings,” the company said. “Rollout starts early July 2026, with no changes to existing compliance or licensing requirements.”

Microsoft also demonstrated what the feature might look like in a Teams meeting, with a screenshot showing a toggle to individually disable the likes of Copilot, Facilitator, and Intelligent recap - or turn off all tools at once.

The company also pointed out its Meeting AI tools will only show up after being cleared by your admins, so specific policy considerations will always be considered - and the toggle will not appear if Meeting AI is specifically turned off by policy.

The news comes shortly after Microsoft revealed Facilitator, a new AI-powered tool which will look to help better manage Teams calls, filling in any potential knowledge gaps which pop up during a meeting.

This has already led some observers to worry about the tool's privacy and security limits, however Microsoft noted it will need to be activated to listen and watch all of your meetings, so it knows when to interfere and chip in.

The toggle is rolling out now, and should complete by mid-July 2026, with Teams users across all devices, including Windows, macOS, mobile, and web, included.

Via WindowsLatest

• Productivity and AI-induced cost reductions are actually dropping
• Most businesses plan to continue spending on AI regardless
• Only 35% have full visibility into costs, reporting lower ROI

Despite ongoing deployment, many organizations are apparently still struggling to achieve ROI from AI, with new KPMG data revealing growing accountability, governance and workforce pressures.

The report found productivity gains actually fell from 42% to 35%, with decision making speed also dropping from 41% to 36%. Even costs were challenged, with cost reductions slightly falling from 31% to 29%.

But planned AI spending data indicates a near-identical value compared with the quarter before, implying companies could be investing blindly without detailed strategies spelling out where they would get the most returns.

AI ROI is still a challenge, years later

Backing that optimism, four in five (79%) say AI would remain a top investment priority even if a recession occurred, with a similar number (78%) confident they can future-proof their AI strategies accordingly.

Costs are clearly being scrutinized, though, with 22% now factoring in lower-cost AI models (compared with 15% previously). Nearly half (49%) have even delayed, paused or shrunk their AI strategies over cost concerns.

"AI is now as much a financial management priority as it is a technology one," Global Head of Advisory Rob Fisher summarized.

Clearly, model capabilities are no longer driving AI investments as companies start to look more closely at how much they're paying for services and the promised returns. And with only one in three (35%) having full visibility into AI operating costs currently, much work needs to be done here.

The report even argues that those with full AI cost visibility are 5x more likely to report ROI. "We’re seeing a clear divide between organizations with leadership accountability at the top and those without," Global Head of AI and Digital Innovation Steve Chase commented.

OpenAI has begun discussions about giving the US government a 5% stake in the company, according to an FT report, with CEO Sam Altman supposedly raising the idea as a method for smoothing relations with the Trump administration.

Of course, right now there is no agreement or deal, and no guarantee the idea will ever move beyond conversations. Any arrangement would almost certainly require political support and significant legal work before it could become reality. Still, the fact that OpenAI is even entertaining the conversation tells us something about how seriously artificial intelligence is now being treated, both in Silicon Valley and in Washington.

The first reaction many people had was understandable. If the government owns part of OpenAI, does that mean ordinary Americans somehow get a share too? It's an appealing thought when AI companies are attracting eye-watering valuations while promising to reshape the economy. Unfortunately, that's not exactly a likely outcome, no matter what the intentions.

AI economy access

The reports suggest Sam Altman has discussed a model inspired by Alaska's Permanent Fund, which invests state oil revenues and distributes annual payments to residents. It's an odd framing of AI as a natural resource instead of a software business. Bullish AI fans insisting it will be economically transformative might see it that way, and if they're right, perhaps some of that value should eventually flow back to the public, many of whom have helped incrementally train the models through use.

But the government owning shares in OpenAI wouldn't automatically translate into everyone getting a check. Financial benefits would depend on lots of little details, including whether profits were distributed at all, and if they'd go to public services or even the national debt over your own bank account.

Despite being just a hint of a rumor of a conversation, the questions are worth taking seriously. AI companies are asking society to embrace changes that could alter workplaces, education, healthcare, and entire industries. It is not unreasonable for people to wonder whether they should share in the wealth created by those changes.

Power at stake

There is another reason these discussions matter, and it may prove even more significant than the financial side. OpenAI has become part of a broader conversation about national economics and technological leadership. Governments around the world increasingly see advanced AI as strategic infrastructure rather than another consumer technology.

That helps explain why OpenAI might want a closer relationship with Washington. AI companies already rely on government decisions. Those connections are likely to become even more important as AI models grow larger and more expensive to build.

But governments are expected to regulate powerful companies fairly and independently. Becoming a shareholder in one of those companies could make that relationship look unethical, even with the best will in the world. Public trust often depends as much on appearances as on legal structures. Especially since there's even less sense that OpenAI's competitors like Google, Anthropic, or Meta will follow suit.

A government stake does not automatically mean the public owns part of OpenAI in any meaningful way, and it certainly does not guarantee anyone will personally benefit. So even if the proposal starts to become more real, skepticism and a close eye on any actual agreements is a healthy approach.

• Claude Science is a new “workbench” to consolidate fragmented research workflows
• Everything from literature review to publication is handled on private infrastructure
• Anthropic continues to roll out industry-specific AI tools for real-world use cases

Anthropic has introduced Claude Science – a new, beta AI workbench it says will let scientists consolidate fragmented research workflows into one unified environment.

With model capabilities no longer holding back AI adoption, the Claude-maker’s solution is to respond to today’s challenges, including limited use cases, struggles deploying AI in real-world environments and difficulties integrating multiple tools.

Claude Science represents this response, packaging existing capabilities into a purpose-built application for life sciences and scientific computing, following earlier work on MCPs, skills and other partnerships. An FAQ on Claude Science’s web page reiterates this: “Claude Science is a public beta app, not a model.”

Scientific ‘workbench’

Anthropic’s clearest message in the announcement is that scientific research is largely held back by workflow fragmentation, not model intelligence, with scientists already juggling tools like PubMed, Jupyter, R, a cluster terminal and more.

“Claude Science brings these fragmented tools into a single research environment where scientists can conduct all stages of their work,” the company summarized.

The platform should help scientists handle everything, from literature review and hypothesis exploration to analysis, figure generation, manuscript drafting and publication.

“Scientific research is inherently visual,” Anthropic wrote, acknowledging that many researchers are being held back in quickly and accurately producing visuals, which could need multiple revisions and finetunes before reaching production.

For full auditability, Claude Science also includes underlying source code, message history and plain-language explanations within AI-generated outputs for scientists to review and audit progress.

“It runs on your lab’s own infrastructure,” Anthropic added, referencing enterprise-grade laptops, Linux boxes or HPC login nodes, “so large or sensitive datasets never have to leave the systems they’re already on, and only the context needed for each step of the analysis is sent to Claude

Science is a growing focus for AI developers

Anthropic says early testers have already used Claude Science for single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction and cheminformatics, by the likes of Manifold Bio, Allen Institute neuroscientist Jérôme Lecoq, and ​​UCSF Brain Tumor Center associate professor and epidemiologist Stephen Francis.

The new tool represents a growing area of interest for AI developers, who are now targeting sectors with industry-specific tools rather than continually upgrading model capabilities without offering clear use cases. Until now, finance and legal have been a major focus for the likes of Anthropic and OpenAI, and this new science-focused initiative could mark the next stage.

It follows rival company OpenAI’s introduction of Prism earlier this year, described as an “AI-native workspace for scientists to write and collaborate on research” that launched with GPT-5.2 – the then-current model.

Claude Science is a separate app that’s available in beta for macOS and Linux installations to Pro, Max, Team and Enterprise subscribers.

The company has also committed up to $30,000 in credits for 50 lucky projects.

• A Chinese robotics firm has said that some workers ‘want to be replaced’ by its humanoid robots
• Agibot has launched its range of humanoid and quadruped robots in the UK B2B sector
• The company's Europe and US lead also said that robots could become nurses and teachers

An executive for Chinese robotics firm Agibot has said the company believes its humanoid robots could replace certain human workers — and that “in some jobs, they want to be replaced.”

Agibot, founded in 2023 by two ex-Huawei engineers, makes humanoid, quadruped, and cleaning robots for the business-to-business (B2B) sector, advertising use cases like manufacturing, cleaning, entertainment, and construction.

The company recently rolled its 15,000th unit off the production line and announced its expansion into the UK market at an event in London on June 30.

In a press conference at the event, attended by TechRadar, the president of AgiBot Europe and America, William Shi, told reporters that “we’ve got to factor people doing dangerous, boring, and repeatable jobs — these kinds of jobs can be very easily replaced.”

Shi added: “For some job descriptions, they want to be replaced, because it’s very boring, very dangerous, and very high-risk — nobody wants to do this [kind of job]”.

‘Replaced by robotics’

Agibot's product range includes the full-sized A3 humanoid robot, the half-sized X2 humanoid robot, and the D1 range of quadruped robots (with form factors resembling dogs, though the company never describes them as such).

Though the company has just landed in the UK, its robots have been successfully deployed in manufacturing facilities in China. A recent YouTube livestream shows the G2 industrial robot ‘at work’ at the Longcheer electronics factory, where the humanoid robots are deployed alongside human employees.

Speaking about Agibot's deployment with Longcheer, Shi said: “They have a lot of workers standing for eight hours per day. They take a smartphone, turn it around, and put it in a box. And then they take the box and move to the next production phase.”

“These kinds of steps can be easily — and are expected to be — replaced by robotics,” he continued, “because that doesn’t create value or create happiness for people. They don’t learn when they do this job. They don’t invent things.”

At present, Agibot's products are not autonomous, but each is equipped with a three-part AI model, with each part controlling interaction (with people), locomotion (moving around), and manipulation (of the local environment — i.e., picking stuff up), respectively.

The company is actively aiming for autonomy, powered by the collection of usage data and further AI development, but Shi emphasizes that Agibot humanoids will remain “under the control and expectation of the human.”

Robot child-rearing

As well as manufacturing, Shi highlighted baristas, live entertainment workers, and even teachers and nurses as professions that could be occupied by robots: “There’s a big lack of nurses in every country, from China, to America, to Europe, and also a big lack of teachers.”

“Most of children’s questions can be responded to [by robots] — phonics, science, and mathematics questions, or even some common conversations. You ask for the weather, the humidity; you can ask these questions to all the robots, because they’re based on large language models.”

Still, there’s some distance between using ChatGPT to conduct research and allowing a humanoid robot to teach children — and according to 2025 research by KPMG and the University of Melbourne, while almost three quarters of UK adults use AI at work, less than half say they trust it.

With the B2C market a distant target for Agibot and an array of business partners — including Nvidia, which provides chipsets for the robots — backing the company's UK launch, it's clearly expecting to make big progress in B2B industries first and foremost. But Agibot isn't ruling out a future in which robots are part of our everyday lives.

“What we want in the future is that the robot can take some responsibility in daily life,” said Shi, “but they will never make the decisions instead of a human.”

Hayao Miyazaki has been at the helm of Studio Ghibli, the legendary Japanese animation studio, in one way or another for 40 years – now serving as honorary chairman and continuing to direct feature films. Over the last few decades, he's seen a dramatic change in culture and technology while sticking to his traditionalist guns.

The death of creativity

Miyazaki is known for aggressively high standards and a dedication to the craft of animation.

For years, he has overseen teams of artists hand-painting beautiful visuals from films such as My Neighbor Totoro and Spirited Away frame by frame. His reaction to learning that artists on his team were experimenting with AI was pretty much as you'd expect.

Filmed speaking to his team during the filming of the 2016 documentary Never-Ending Man: Hayao Miyazaki, the animator was unequivocal in his condemnation of the use of AI for an experiment during the production of the short film Boro the Caterpillar.

Rage against the machine

Miyazaki shared his thoughts in this documentary a year before Google released the seminal paper on the transformer architecture that now underpins today's dominant AI models – and six years before OpenAI first launched ChatGPT.

Despite airing his views so bluntly, nearly a decade on, OpenAI's image generator became embroiled in controversy when thousands of people took part in a viral craze that involved generating Ghibli-esque graphics within seconds.

The trend raised serious concerns about the role that AI plays in the modern creative landscape. While it was used primarily for fun in this instance, it's easy to see how executives may see the time-saving potential of automation tools and the allure of removing the blood, sweat, and toil that humans put into their creative endeavors.

It seems that Microsoft explored the idea of building Windows fully around AI in the past, based on a leaked video from a couple of years ago.

Windows Central highlighted a video (see below) that's a few minutes long and was leaked via the BetaWiki Discord server, with our sister site's Zac Bowden noting that sources have provided assurances that the clip is real. It shows an AI-focused version of Windows built around Copilot and apparently codenamed Aion.

The concept shown is a lightweight web-based OS, meaning it's built on web apps rather than native Windows apps. In other words, it won't run standard Windows (Win32) software, with the idea being to stream those apps to the desktop if they're required (meaning they're run from the cloud, or more specifically, Windows 365, Microsoft's cloud PC offering).

It's kind of like Microsoft's take on ChromeOS, then, leveraging the cloud, except that it's built around the Edge browser and Copilot.

Copilot runs the show, and is the central player in the Start menu, and the idea is that AI provides contextual suggestions here, recalling previous interactions to try and anticipate what the user might need.

In the video, Microsoft explains that Aion aims to break down the "traditional app-centric" approach to grouping on the taskbar, instead using 'Spaces' that act as groups into which apps, websites or files that pertain to the same goals are deposited.

Analysis: AI-on or AI-off? It seems most lean towards the latter

The themed approach of Spaces sounds rather like the idea of Sets that Microsoft toyed with in Windows 10 the best part of a decade ago now, only to abandon the concept. Except this time around it's grouped content that's organized and curated by AI.

The Aion concept hasn't been well received by the computing public as you might guess. One commenter on the video simply states: "This company has completely lost the plot."

Another observes that it's "like ChromeOS for people who don't know how to use a computer at all."

And yet another notes: "How did they manage to even make simple web apps look slow and laggy? One of the strong points of ChromeOS is that it is very fast even on old, slow machines."

In fact, there are a few people who aren't impressed with how clunky and sluggish the operating system appears to be in the video. In fairness to Microsoft, though, it's just a concept illustration and early working code (although the evident lack of smoothness isn't a good look, it must be said). Bowden explains that the video was recorded at some point in 2024, and that it's "unclear if this was just a Hackathon project or something more."

The ideas explored within Aion could well be a hint of where Microsoft is headed with next-gen Windows, though. Which may be worrying for some, of course, but you might as well get used to these ideas.

While Microsoft has promised to trim back AI excesses in Windows 11, that's more about streamlining submenus here and there, and removing Copilot features from certain apps, than it is some kind of wholesale change of philosophy regarding AI. Windows 11 is getting AI agents, and indeed they are the next big thing for the OS, if Microsoft has anything to do with it (and, strangely enough, it does).

Indeed, with Project Solara, Microsoft plans to bring AI agents to all manner of devices in the world, beyond mere PCs and phones. Bowden theorizes that maybe Aion evolved into Solara.

Whatever the case, Aion is still a thing, believe it or not: Microsoft revealed a new family of local AI models running with the same name at Build 2026. These are a "new generation of small language models that are smaller, faster, and more efficient than our previous Windows OS SLMs", as Microsoft explains here. Apparently, Aion lives on in some form, then, even if it's a very different idea to the notion of a full-on Copilot-based operating system.

• SAP is cutting travel spend on refocusing on AI-related hiring
• Company shares are down 46% in 12 months over software concerns
• Redeploying existing workers is a priority over laying off more employees

SAP is reportedly revising how it spends money to free up more money for its AI strategy, with a new internal memo seen by Bloomberg seemingly confirming the company's intentions to restrict new hiring, pause internal travel and reduce other spend related to suppliers.

It's believed the company will still continue hiring across certain AI roles, showing a shift toward AI engineers, researchers and other specialists, but other roles will likely see a slowdown or pause.

The company told employees that AI is reshaping enterprise software, so targeting investment to keep it competitive long-term would be crucial.

SAP redirects internal spend to focus on AI

According to the report, citing an internal email, company travel unrelated to AI projects and customer relations has been suspended. The company also wants to redeploy existing workers to plug new gaps, rather than firing and rehiring. SAP laid off around 12,000 workers between 2023 and 2024.

"We are prioritizing investments in AI-related capabilities, talent, and technologies while applying greater discipline to hiring, external spending, and internal travel," a company spokesperson said.

"I don’t expect to operate with a smaller work force, but with a very, very different work force," CEO Christian Klein had previously told the New York Time (via Investing.com), implying that jobs would continue to evolve.

Despite posting a 6% increase in its most recent quarter's revenue, SAP shares have dropped around 46% over the past 12 months amid concerns that its software business might face longevity struggles amid the ongoing AI boom.

• Companies going all-in on AI are also seeing positive impacts on headcount
• Entry-level roles saw higher-than-average growth, opposing existing research
• Workforce gains aren't immediate – AI needs time to find its place within organizations

A new report has challenged the narrative that AI adoption has been leading to job losses, instead revealing that the companies making the biggest investments in AI are actually growing their workforces.

The study combines corporate AI spending data from Ramp's payment platform and workforce records from Revelio Labs to analyze more than 21,500 US companies, making it one of the biggest of its kind.

It concludes that high-intensity AI adopters increased their headcount by around 10% during the first two years after deploying AI, making AI good news for workers and labor after all.

AI adoption is causing companies to hire more workers

Clearly, only strong AI adoption has a positive impact on workers, because companies making modest investments didn't see any significant growth.

The study also stresses that the impacts are slow-growing – rather than seeing an immediate uptick in employment, it takes time for companies to integrate AI, discover productive use cases and hire more workers.

High adopters are defined as those who invested around $33 per employee per month during the first three months after adoption, compared with around $3 for low adopters.

It also challenges other recent research, asserting that entry-level employment actually rose by a higher-than-average 12% among high-intensity AI adopters. Other reports have implied that entry-level workers have been among the hardest hit.

Even though tech giants dominate the headlines, with Salesforce cutting nearly half of its support staff and Amazon notably cutting tens of thousands of workers, the Ramp/Revelio Labs report actually shows growth across more than just AI engineer roles, spanning sales, marketing, admin, finance, customer service and more.

While the research can't be used to predict long-term labor impacts, it does at least serve as a notice that workers aren't currently at risk of total redundancy, even amid job role shifts.

• KnowBe4 finds 55% of UK workers admit to using unapproved AI tools
• Only 16% believe they're effectively managing safe AI
• 27% supplement corporate AI with more suitable tools

More than half (55%) of UK employees admit to using unapproved AI tools at work, with as many as one in 10 knowingly sharing sensitive company information with these unauthorized tools, new research has warned.

The new report from KnowBe4 serves a purpose of defining shadow AI as unapproved AI, not AI use that goes under the radar, because the number of UK cybersecurity decision-makers identifying shadow AI as the biggest risk is nearly matched (58%), implying they're well aware of the challenges.

However, little seems to be being done, because only 16% believe their organization is effective at managing AI's safe use at the moment.

Shadow AI is an ongoing challenge

Nearly half (46%) have implemented targets to improve AI agent safety over the next 12 months, but with one in five (19%) already reporting that AI agents take autonomous action across multiple workflows with limited human oversight, the risks remain clear.

"UK businesses are embracing AI to drive productivity [but] many employees are still under pressure, using unapproved tools and regularly facing (and fearing) sophisticated threats such as deepfakes and phishing," Lead CISO Javvad Malik wrote.

Shadow AI doesn't necessarily mean that employees are turning their backs on enterprise-grade tools, with 27% admitting to occasionally sourcing their own tools on top of the tools they're given, indicating that companies are failing to provide the right tools that workers need.

So while governing AI's use with clearly defined policies is one area for improvement, simply providing workers with the tools they're demanding could go a long way to reducing shadow AI's impact on any organization.

• Dassault Systèmes secures £30m deal to support UK fusion energy program
• 3DEXPERIENCE platform will enable digital twin modelling, testing and more
• STEP program could create 10k+ jobs – including Dassault opportunities

Dassault Systèmes has teamed up with UK Fusion Energy in a new £30 million deal to support the Spherical Tokamak for Energy Production (STEP) programme, supporting the UK’s effort to build the world’s first prototype fusion power plant by 2040.

Under the new contract, Dassault with provide the project with its 3DEXPERIENCE platform to support digital twins, crucial for developing next-generation power plants and broader fusion research.

The software will also be used for engineers, designers, scientists, manufacturers, construction partners and more to collaborate with the same data and context, ultimately creating a hub for the project. Dassault described the tool as a “standardized and connected engineering environment.”

Dassault Système’s role in digital twins

The multimillion-pound software package reduces the need for separate CAD files, simulation models, engineering databases and more, improving the project’s overall value.

It will also be used for engineers to model, test and optimize systems using digital twins – a digital mockup of the power plant in a virtual environment.

“[STEP’s Plant Information Management System], powered by the 3DEXPERIENCE platform, will play a critical role in enabling fast, efficient and rigorous engineering delivery while creating the blueprint for future fusion power plants,” UK Fusion Energy Engineering Program Director Debbie Kempton wrote.

“This deployment will strengthen engineering continuity, accelerate innovation and ensure that plant information, engineering decisions and system designs remain fully connected across the entire lifecycle of the prototype power plant,” Dassault Systèmes EuroNorth MD John Turnbull added.

STEP, which was awarded a £2.5 billion of additional government investment in June 2025, is set to be built on the site of a former coal power station in Nottinghamshire.

Fusion power plant inches toward reality

The announcement also marks the next step in this drawn-out program, moving it beyond scientific research into pre-production.

Already this year, Tokamak Energy has signed a £70 million contract to develop high-temperature, superconducting magnets that will fuel the fusion power plant.

The government anticipates the project could create up to 10,000 jobs across construction, engineering and supply chains, including the impact the new £30 million contract will have on employment at Dassault Systèmes. It’s unclear how long the deal will last, and whether Dassault will need to re-tender in order to welcome competition.

Final designs should be complete with construction expected to start around 2032, with around eight years of construction, installation and other commissioning to take the site live.

As for Dassault itself, the £30 million contract represents a major contract for the company, which scored another major deal with the Volkswagen Group in early 2025 to support thousands of engineers with virtual twins throughout vehicle development across brands like VW, Audi and Porsche.

• Report claims degree holders and tech sector workers continue to see AI-induced layoffs
• New unemployment tracker lays foundation for future employment laws
• Despite layoffs, new job opportunities remain elsewhere

A New California Policy Lab report has revealed artificial intelligence could actually be leading to a number of layoffs despite recent promises that the tech is creating new roles.

But the data points towards certain roles being more susceptible than others. Those with a bachelor's, master's or PhD degree in highly AI-exposed roles saw higher unemployment rates after ChatGPT was released in late 2022.

Insurance claims for unemployment among bachelor's degree holders rose more than 50% between November 2023 (13,000 claims per month) and July 2023 (22,000 claims per month). Claims have since fallen, but remain above earlier levels, standing at around 16,000 claims per month currently.

AI still causing some job losses

And California leads the way for AI-induced layoffs, with the Bay Area's high concentration of tech companies making it a likely place for labor market effects to emerge first.

But the report is largely optimistic about AI's impacts on roles – although certain roles are clearly at risk of displacement, other roles still promise opportunities for human workers.

"AI does not seem to have affected national unemployment rates... but has affected the headcounts of certain occupations with more exposure to AI," the paper concludes.

Though region-specific, the report marks the introduction of the California AI-Unemployment Tracker (CAIT) using near-real-time data, which it says is important for guiding employment laws and support as AI's impacts on the labor market continue to evolve.

Ultimately, the paper concludes that higher-level workers are being affected first, particularly in Califorina's tech sector, but the overall impact on the labor market might be more displacement rather than replacement.

Anthropic has just released Claude Sonnet 5 for all users, and I wanted to test what it was good at. But the game has changed now. Sonnet 5 doesn't feel dramatically different from Gemini or ChatGPT if you ask it ordinary chatbot questions. Instead, the difference should show up when you stop asking for answers and start asking for completed work.

Anthropic says Sonnet 5 is built for "multi-step software engineering work," sustained coding, tool use, debugging, and "messy technical contexts." It also says it can make plans, use browsers and terminals, and run more autonomously than smaller, cheaper models previously could.

I'm not using Sonnet 5 for coding, but that doesn't mean I can't take advantage of its new abilities — just like you can. So I stopped asking Claude for answers and started asking it to finish jobs, beginning with planning a trip to Bath, UK, for my family: my wife, me, and two teens.

A trip to Bath

When I tested it, Claude Sonnet 5 defaulted to its Medium level of effort, so that's what I used. Here's the first prompt I tried:

"I want to test whether you can act more like an agent than a chatbot.

My task is: Plan a weekend trip to Bath for two adults and two teenagers, including travel, lunch, one activity, estimated costs, and what still needs booking.

Don't just give me advice. First, make a brief plan. Then identify which parts of the task you can complete yourself right now, which parts require tools or information you don't have, and which parts need human judgment.

Then complete as much of the task as possible without stopping after the first obvious answer.

At the end, give me:

What you completed

What still needs human action

Any assumptions you made

A short checklist I can use to verify the result

The next best step"

What I really liked was that, as Claude tackled this task, it gave me the option to be notified when it had finished. In reality, it only took a few seconds to come back with a plan, which included travel options, an itinerary, and a suggestion for lunch and something to do: a trip to The Roman Baths.

To my delight Claude gave me an interactive map showing where all the places it recommended were. It also gave me a useful list of what it had completed, what required human action, the assumptions it had made, a verification checklist, and a "next best step" action point. It felt ready to keep working with me as more details came in, rather than treating its first answer as final.

In fact, when I gave it more details, such as which day I was going to go, it gave me a visual weather report for the day. That was a really nice touch.

Claude vs ChatGPT

I also tried this prompt with ChatGPT-5.5 Medium and got a similar result. It acted as an agent, just like Claude did, and notified me when it had finished its tasks. It just didn't look as nice. There was no map, or any visual elements at all, and it felt more like I had been given a finished report than the start of a two-way conversation where it asked me for more details.

Both chatbots recommended lunch and a trip to The Roman Baths. Interestingly, ChatGPT assumed I’d get the train, while Claude assumed I’d drive. They also recommended different places to eat, but the core information they both provided was solid.

What was most impressive was that both models could adapt when I reframed the inputs. For example, when I gave them the ages of the kids, student status, a different mode of transport, or changed the day of the trip, both models could cope. Both also identified that since the oldest was a university student, he could get free entry to The Roman Baths.

This part of the test was probably the most meaningful, as it felt much more "multi-step" than simply providing one answer.

Overall, I’d give this test to Claude. You can clearly see that Sonnet 5 is set up for agentic actions. Neither Claude nor ChatGPT could actually do any of the booking for me at the moment, so we're still a long way from true personal-assistant-level autonomy. But for this kind of task, Claude currently has the edge.

A different domain

I wanted to test the models in a different domain that would let Claude show me it had genuinely improved, and that the Bath trip result was not just a fluke of the travel-planning use case. So I asked them both to:

"Build me a simple household budget tracker as a spreadsheet or small tool."

Both models thought for a while about this task, and churned through various options before opting to make a spreadsheet. ChatGPT produced a spreadsheet with a bar chart that tracked how much I’d spent on various household expenses against a budget. Claude, however, went for something simpler: dispensing with a budget, it just tracked actual expenses and created a pie chart showing where my money was going.

Claude’s initial approach was simpler, and easier to understand. Both models provided a .xlsx file, but only Claude provided a button to upload it straight to Google Drive so I could open it in Sheets.

I told ChatGPT, "I wanted the graph to be a pie chart," and it responded: "Absolutely — I’ll update the spreadsheet itself so the dashboard uses a pie chart for spending by category, rather than the current graph style."

It ran into a few problems because it was trying to show both the budget and actual values in the same pie chart, but eventually it worked out that it could show only one and produced a new spreadsheet that did exactly what I asked for.

I then asked Claude to change its spreadsheet to provide a budget section too, and to change the graph into a bar chart. Again, it showed me its workings and added a budget section and bar charts perfectly.

I can’t really separate the two AI models on this task. Both proved they can handle multi-step tasks well, and both were happy to revise the result when I changed the brief.

That, really, is the point. The most interesting AI tests now are not "which chatbot gives the best answer?" They are "which assistant keeps working until the job is actually done?"

On that front, Claude Sonnet 5 feels extremely capable. ChatGPT was close behind, and in some ways just as effective, but Claude felt more naturally organized around the idea of completing work rather than simply responding to prompts. It asked fewer invisible questions, presented its output more helpfully, and made the whole process feel more like collaborating with an assistant than interrogating a chatbot.

For now, neither model is ready to fully take over the job. I still had to check the details, make the decisions, and do the actual booking or uploading myself. But the direction of travel is obvious. The AI war is no longer just about who has the smartest chatbot. It’s about who can build the assistant that gets you closest to a finished task.

Google's new Nano Banana 2 Lite is really fast, as in about four seconds from prompt to image. That speed changes how you think about writing the prompts as much as the schedule of producing them.

The standard Nano Banana 2 model, and most other AI image generators for that matter, take long enough that it's worth spending some time working out the perfect prompt. It can be annoying to have to redo it multiple times when you have to wait up to a minute and still might get it wrong. You learn to be cautious in your prompting.

Nano Banana 2 Lite breaks that rhythm. I noticed my own speed changing to almost match. I didn't feel the need to write a perfect prompt. I treated it more like a sketchpad for ideas that could be quickly tossed out if they didn't work or revised until they did. None of them felt like much of a commitment because another attempt was only a few seconds away.

And it's not as though there's an enormous downgrade in results. For instance, I asked both versions of Nano Banana 2 to make "A steampunk fleet sailing through outer space above Earth, complete with ornate wooden airships covered in brass." There is plenty there to cause fits in any image model.

Without knowing which was which, more than one person guessed wrong or thought it was a trick and the two were from the same model. The one on the left is Nano Banana 2, and the one on the right is its Lite sibling. You can guess one is higher quality if you study it, and certainly over time you can spot where the Lite version might let you down, but when it only takes four seconds to come up with another one, it doesn't matter too much.

Speedy creation

While the standard Nano Banana 2 is good for when you need the highest fidelity or have an extra tricky request, Lite exists for speed and brainstorming. Google positions Nano Banana 2 Lite as the faster, cheaper companion, helpful especially at scale.

For the average person, it means you don't have to invest as much time in your initial prompt and can play around more.

It means you can iterate in interesting ways too. For instance, I started with a prompt for "a busy farmer's market." Crowds remain one of the quickest ways to expose the weaknesses of AI image generators because there are so many people, poses, and interactions happening at once.

The result was fine, but I began adding specific details and Nano Banana 2 Lite obliged me with about a dozen options in a few minutes. Now, my requests for things like children chasing bubbles, an elderly couple buying flowers, a street musician, and a fruit vendor making a sale in the foreground are all right there. There are some flaws and odd details, but for four seconds it's not bad.

Brainstorming prompts

And it should be repeated that Nano Banana 2 Lite is not much of a step down from the larger model and is capable of coherent storytelling, but not necessarily perfectly on the first try. I asked it to "Make a six-panel comic about a businessman who accidentally swaps briefcases with an alien in a train station."

The comic makes sense overall. The characters stayed consistent from one panel to the next, the sequence flowed naturally, and the final reveal landed with exactly the right amount of absurdity. There are, of course, two huge errors in the comic, where the human has no briefcase in the second panel and two in the third.

A couple of further prompts solved the problem, but it's important to note that the easier solution is to iterate, not to spend a lot more time reworking the prompt with extra detail. A version made with the regular Nano Banana 2 model notably did not share the same flaw. In other words, each model has its place, and you might even end up polishing a prompt in Lite and then taking it to the Nano Banana 2 for an even higher-quality version.

Still, when you can make so many images so quickly, you can rethink how you come up with the prompts. That seems to be Google's goal for Nano Banana 2 Lite. It feels designed for the messier parts of the creative process. The bigger model is for a deeper commitment. Making each individual image feel a little less important might eventually encourage people to create far better images.

Palantir CEO Alex Karp has hit out at the state of the AI industry, saying it was "effing insane" that the technology is being used in areas such as military and national security.

In a heated interview with CNBC Squawk Box, the controversial billionaire also hit out at top AI firms such as OpenAI, claiming he had spoken to major CEOs outside of the industry who were "livid" at how some companies are doing business.

Karp also accused some major AI companies of imposing a “wealth tax” on businesses by charging high fees for their services, all while collecting data which may be used to improve their own AI models and tools.

"Completely wrong"

Karp's ire was particularly focused on the token model being used by the likes of Anthropic and OpenAI, especially as costs continue to rise, but companies look for a better return on their investment.

“I’m not throwing shade at them, but something has gone completely wrong,” he said. “The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens.”

This includes a range of Chinese firms, with Karp warning the US not to underestimate the speed of progress being seen at its great rival.

Rising AI prices have led many businesses to pivot towards building and training their own models, rather than relying on outside providers, with so-called "open weight" models able to perform at a fraction of the cost.

Karp's frustration was clearly visible, with one CNBC host commenting, “You sound pretty angry,” with the CEO responding, “This is the voice of American business that is being channeled through me.” 

To shore up its own support, Palantir recently announced a major partnership with Nvidia which will see the latter's AI services used to create custom models for US government agencies.

“What aligns me with Nvidia, and I think is what the technical customers want, which is control over their compute, their models, their data stack and their alpha,” Karp told CNBC. “They want to know they own the means of production. It’s not being transferred to someone else.”

This follows recent criticism by the US government of firms such as Anthropic, whose Mythos 5 and Fable 5 AI models were deemed a national security risk and shut down shortly after release.

Karp went on to criticize the US government for its reliance on AI companies in creating new technology for the military and national security.

“Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane," he noted.

• UBTech just introduced its most-human-yet robot
• The silicone skin is realistic and creepy
• It launches in China this year, but there are also plans to customize it to look however you wish

Ubtech, last seen deploying humanoid robots to patrol the China-Vietnam border, is back with its creepiest creation yet, the UWorld U1, its first mass-produced, full-sized "ultra-bionic humanoid robot."

I know that's a mouthful, but the UWorld U1 Series does mark a departure for UBTech, which has traditionally produced faceless automatons aimed primarily at business, enterprise, and the border. However, the company has long had consumer ambitions, and the UWorld U1 Series may be the bot to realize them.

This new humanoid robot features silicon skin, a lifelike face with eyes that follow you, lashes that blink coquettishly (ick), and 88 degrees of freedom across its full-sized robotic body. It even has what UBTech calls a "dual-pivot biomimetic cervical spine," which apparently gives it more human-like movements.

UBTech unveiled the robots earlier this week at a splashy launch event in Shenzhen, China. There are three models in the Series: the U1 Lite semi-torso edition, the high-performance full-body U1 Pro, and the high-dynamic full-body U1 Ultra.

During the launch on June 30, 2026, several UWorld U1 Ultra humanoids walked the stage, interspersed with actual humans. The idea, I think, was to confuse the viewer so they might not know which was which. In truth, the robots all looked a little plastic, or like characters straight out of your favorite anime, and they walked somewhat awkwardly.

To further prove their believability, UBTech had one robot dance with a tuxedoed human. At times, it looked as if he might be supporting a potentially teetering UWorld U1 robot.

What are these robots for?

In addition to questionable dance moves, UWorld U1 is equipped with an "emotion-aware LLM" that the company claims will help it recognize and respond to "fine-grained emotional states." It's designed as a proactive companion, responding to human interaction in a fraction of a second.

UBTech also, somewhat comically, promises a focus on privacy, in a country where the Chinese government can ask to see all your data at any time. Still, UBTech's privacy architecture is smartly focusing on "minimal cloud dependency," which means most of your data may not be with UBTech anyway.

According to a release, the UWorld U1 is designed to meet a real and growing need in China, where the company claims that 90 million adults are living alone and there are 118 million empty-nest senior citizens.

The company is so concerned about this growing issue that it plans to donate 100 of these robots in 2026. But here's where things get really weird. I'll let Ubtech speak for itself:

"These units will incorporate 3D facial reconstruction and voiceprint-based identity replication technologies to recreate designated individuals, while integrating emotion-driven interaction models and dedicated long-term memory systems."

As I read that, they plan to make these robots look like the person of your choice. Perhaps it's a dead husband or maybe an adult child who moved away. The custom UWorld U1 Ultra will look, sound, and maybe even respond like your missing companion.

Yikes! How did we get here? Why is life imitating Black Mirror? In truth, these robots are unlikely to fool anyone into thinking they're really human. Sure, UBTech did its best to confuse us in the presentation, but in the real world, that cold silicon, odd gait, and sure-to-be-creepy interactions won't be fooling anyone and will be a poor substitute for your absent partner.

Also, if you really are interested, you'll have to move to China and plop down the equivalent of almost $18,000 in US dollars (around £13,570 / AU$26,120) when they ship sometime this year.

ChatGPT's new finance feature lets the AI chatbot take a look at any bank or similar accounts you care to open up for inspection. I was initially hesitant to try it out, but the tool only looks at the details of how you spend your money, and can't actually carry out transactions, so I agreed to let it analyze some of my accounts and offer its insights.

Finances is currently only available in the U.S. to Plus and Pro users on web, iOS, and Android. Setting everything up is as easy as using any other ChatGPT plug-in. You just select Finances in ChatGPT and then click Get Started then Connect with Plaid.

Finances uses Plaid to link to the accounts; you simply sign in and agree to let Plaid share the information. The service synchronized recent transactions and basically became an interactive, conversational database of my information. It's not that the AI can do things it couldn't before, but now you don't have to manually enter any finance details you want to ask it about.

ChatGPT looks at the receipts

At its suggestion, I asked ChatGPT, "How much did I spend eating out this year?" ChatGPT soon had a breakdown of my how much I had spent at various restaurants, including delivery, and broke down some of the patterns. It pointed out that it had gone up recently, which makes sense as the weather has became pleasant enough to walk into town more often to eat. It also identified a few spikes I could tell were around birthdays and family celebrations.

Next I asked, "What subscriptions am I paying for?"

That turned into one of the most practical moments of the experiment. ChatGPT gathered recurring charges across streaming services, software, and other monthly payments into one clear summary instead of leaving me hunting through statements. It also suggested reviewing whether every subscription still deserved a place in my budget.

"These subscriptions appear consistently each month, making them reliable recurring expenses.", it told me.

Financial chatter

Seeking a more detailed analysis, I asked ChatGPT, "What's changed in my spending this month?"

Instead of overwhelming me with every purchase, ChatGPT focused on the transactions that genuinely stood out. It noted a few large purchases at new locations, and asked if they were something I wanted more information on. I told it those were gifts, and the AI incorporated that fact into its response as not a new regular event but one that explained why overall spending had climbed this month.

"Those gift purchases are unusual compared with your recent history but they account for much of the increase, so your everyday spending has remained relatively consistent despite the higher overall total. Looking beyond one unusual month provides a more accurate picture of your longer term financial habits."

Using ChatGPT this way felt almost like the inverse of most financial software. Most budgeting apps expect you to learn their system first. ChatGPT let me ask questions however they occurred to me and then translated months of transactions into answers that were easy to understand.

That does not mean it replaces dedicated budgeting tools, and it certainly cannot make smarter financial decisions on your behalf. You still have to decide whether to cancel subscriptions, spend less on eating out, or save more each month. But I can see how it might make it easier for people who hesitate to look at their finances if they feel uncomfortable around spreadsheets. It didn't feel like math homework the way it often does.

Of course, it still relies on some trust in both OpenAI and Plaid, but as long as it's purely viewing and not actually touching the accounts, this could be a really useful, practical feature for ChatGPT users.

• Anthropic has released Claude Sonnet 5, calling it its “most agentic Sonnet model yet”
• The new model is designed to make plans, use tools like browsers and terminals, and run more autonomously
• Sonnet 5 is available across Anthropic plans, and is now the default model for Claude Free and Pro users

Anthropic has released a new version of Claude, called Sonnet 5, which it’s calling “the most agentic Sonnet model yet.” Agentic models are designed to do more than simply answer questions. They can plan, use tools, and carry out tasks with less step-by-step input from the user.

According to Anthropic, the new Sonnet 5 can “make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.”

Sonnet 5 is aimed at coding and everyday professional work. Anthropic says the latest version outperforms the previous Sonnet 4.6, scoring 80.5% in Agentic Coding using Terminal-bench 2.1, compared to 67% for Sonnet 4.6.

Despite being aimed at professionals, the new release isn’t being restricted to paid users. It's available across Anthropic plans and is the new default model for Free and Pro users, and is available to Max, Team, and Enterprise users as well. It’s also available in Claude Code and on the Claude Platform.

The age of the agents

The release of Sonnet 5 marks a wider shift in the AI race. Chatbots are no longer just competing to sound smarter in a conversation. They are increasingly competing to act like agents — tools that can plan, code, browse, investigate problems, and complete work with less hand-holding.

Sonnet 5 arrives soon after the release of Gemini Spark, Google’s 24/7 agentic personal assistant AI.

Anthropic is also launching Sonnet 5 at the same time that Fable 5 and Mythos 5 have become wrapped up in government scrutiny. Claude Fable 5 has just been re-released after being restricted by the US government, while OpenAI’s GPT-5.6 is still under review.

How AI models are changing

Claude Sonnet 5 may look like just another model launch, but it points to a bigger change in how AI companies are competing. The next stage of the AI war will not be won by the chatbot that gives the neatest answer. It will be won by the assistant that can take a messy task, keep track of the plan, and actually get something useful done.

AI assistants will increasingly complete tasks rather than just suggest steps. In this new future, the best model may not be the one with the cleverest answer, but the one that can finish the job.

• AI is forcing Apple to deliver security updates more often
• iOS 26.5.2 is part of Apple's new update strategy
• More updates mean better protection against AI-powered cyberattacks

It seems that the number of ways AI is changing the world is increasing. The requirement for more RAM to run new AI features in Apple products is already being blamed for the recent price increase in Apple products, as well as for the current high price of RAM itself.

Now, according to a recent Reuters report, AI is also to blame for the number of iOS and macOS updates we'll need to install. The bad news is that it's going up, all because of the threat posed by the latest AI models and their potential to aid cyberattacks.

Instead of waiting for the next scheduled operating system update for the latest round of security fixes to arrive, Apple is now delivering individual security updates ahead of the next iOS and macOS 26.6 update.

iOS 26.5.2 is here now

If you look in Settings > General > Software Update on your iPhone, you'll see that iOS 26.5.2 is waiting for you now, unless your iPhone already installed it overnight.

The description for the update reads, rather vaguely: "This update provides security fixes for your iPhone," but it's the dangers posed by AI that are driving this update, so make sure you install it promptly.

Details of the security updates for all Apple operating systems are available on the Apple Security Updates page.

Malicious hacking tools

The Reuters article states that the urgency of the update is due to the risk that AI now poses to Apple devices.

"The company told Reuters on Monday it was adapting to the reality that, given the ability of artificial intelligence to speed the development of malicious hacking tools, it needed to reduce the time between when updates were first made public and when they were put into customers' hands."

It looks increasingly likely that this will become the new normal for security updates, and companies like Apple won't be able to rely on bundling the latest security fixes into the next scheduled operating system update. Stand-alone security patches delivered whenever they're needed are something we're all going to have to get used to.

While it might be annoying to keep updating our tech gadgets more often, it's a small price to pay for better security in the AI era. We'll just have to get used to software updates being less about new features and more about staying ahead of increasingly sophisticated cyberattacks.

• Bipartisan bill wants big tech companies to pay for the additional energy consumption and capacity fueled by data centers and AI demand
• Congress will begin consideration of the bill, before it moves on
• Electricity prices have risen dramatically in the US, primarily because of an explosion in new data center construction projects

A bipartisan bill is currently moving through Congress that seeks to force the big tech companies behind the massive buildout of AI datacenters to pay for the energy they use.

Currently the cost of energy is calculated by a combination of factors such as how much it costs to generate and transport the power, with extra charges added for plant and infrastructure maintenance, local regulations, and additional capacity and demand charges.

That means those living in areas with data centers connected to the local grid will have to bear the cost of the sudden increase in demand and capacity, even if they are not using more energy themselves. The Ratepayer Protection Act looks to shift that burden onto the tech companies.

Ratepayer Protection Act

The name may sound familiar, and that's because the bill essentially looks to codify parts of Trump’s Ratepayer Protection Pledge — the commitment that companies such as Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed back in March this year. This pledge was only voluntary though, and lacked any real detail on exactly what signatories would be paying for.

The House of Representatives will now begin consideration of the bill, which seeks the introduction of a “large load standard” requiring tech companies to bear the cost of the energy they use, alongside upgrades to the local grids they connect to. The bipartisan bill’s primary sponsors are Reps. Gabe Evans, R-Colo., and Kathy Castor, D-Fla.

“Families and small businesses across the country shouldn’t be left to foot the bill for this new development, though the benefits of these innovations will be felt by all of society,” said House Energy and Commerce Chair Brett Guthrie, R-Ky. “The Ratepayer Protection Act is a bipartisan effort, which would ensure that the costs of grid upgrades are appropriately paid for according to demand.”

If you weren’t already aware, there has been a nationwide wave of opposition to data centers, with both local grassroots movements and nationwide opposition groups successfully delaying and cancelling data center construction projects. There have been a range of motivations behind opposition, from ecological conservation and energy prices, to the fear of AI-influenced job losses and a modern resurgence of the Luddite movement.

The Ratepayer Protection Act is therefore a response to the growing resentment experienced across the US political spectrum, and represents one of the first Congressional attempts to force big tech to pay their way for the enormous increase in energy demand the US has experienced in the past few years — primarily driven by AI demand.

Speaking on behalf of his constituents, bill sponsor Evans said, “Colorado families, farmers, and small businesses should not be forced to cover the costs of new power generation driven by these developments.”

After passing through Congress, the bill will have to be considered by the Energy and Commerce Committee, the House, and the Senate before being ratified by President Trump. Whether the bill will still have teeth in that time remains to be seen. But the bill comes at a pivotal time in the US election cycle, with midterm elections just months away and data centers being a serious consideration for any sitting representative or hopeful candidate.

• Norton's scam detection tools are now available in Claude and ChatGPT
• Users can ask their preferred AI chatbot about the legitimacy of an email, text, website
• Most threats consumers face now come from scams, phishing and fake ads

Claude is the latest AI assistant to get access to Norton's Genie scam detection tool following its available for ChatGPT customers earlier this year.

Available across all Claude subscription tiers, Genie gives users access to scam detection capabilities and other cyber safety tips and advice.

Norton says its tool can analyze suspicious emails, texts, messages, images and links using its "multi-layered" detection intelligence.

Norton scam detection now available in Claude, ChatGPT

"AI assistants are becoming part of how people make decisions and evaluate information online," Head of Products and Portfolios Travis Witteveen noted, hinting that the increased prevalence of AI assistants.

"By bringing Norton Genie into even more AI platforms like Claude and ChatGPT, we’re making trusted Cyber Safety intelligence available directly in those moments to help people make more confident decisions in real time."

The company explained that Genie looks for language patterns, social engineering tactics, urgency cues, impersonation attempts, and requests for sensitive information. It also checks URLs and analyzes domains to confirm whether a user should click on the link.

When the tool launched for ChatGPT in March 2026, Norton described it as the "world's first AI-powered scam detector." Users can start conversations by tagging @Norton and asking questions like whether an email looks legit or if an online offer looks like a scam.

The company's own reporting reveals that nine in 10 threats targeting people in 2025 came from scams, phishing and fake advertisements.

So far, Norton looks to be the only security company offering direct AI chatbot integration to provide accurate insights into threat detection.