While we didn't quite get another 12 days of OpenAI news like last year, the company behind ChatGPT did have a quick live stream on March 25, 2025.
The news? Well, while the AI behemoth was tight-lipped in the lead-up, OpenAI did debut native image generation for the 4o model.
It made the teaser image of someone writing “Livestream at 11AM PT” on a classic, dark green chalkboard make a lot more sense.
OpenAI's much-improved image generation skills are debuting shortly after Google added native image generation to Gemini inside its AI Studio.
Quite possibly the best news, though, is that OpenAI isn't wasting time with the rollout. During the stream, it started rolling out the features to ChatGPT users, and native image generation for the 4o model is available now for all, regardless of your plan. Pro and Plus subscribers get more access, as you might suspect, as folks on the free plan will deal with some limits.
In our early testing, the quality of the images requested was certainly improved compared to the previous Dall-E model, but these also took longer to create. The latter is something that OpenAI called out during the live stream, but it could also be that the company is ramping up resources to handle the demand right after the launch.
Ahead, you can see TechRadar's live blog during the event as OpenAI CEO Sam Altman walked us through the news and updates, before we put the new feature to the test.
Well, the livestream title is shedding a lot more light as to watch we can expect ... way more than the intitally teased image. It's titled "4o Image Generation in ChatGPT and Sora" so that means we're likely getting improvements to creating images within ChatGPT and Sora.
The mention of the latter might mean more general improvements for text-to-video generation as well.
Under 15 minutes to go now!
OpenAI's live stream has begun, and in the lead-up to the 2PM ET / 11AM PT / 6PM GMT start time, we're being treated to various images. Some of these overlap, but it refreshes every few seconds and shows off all the different styles.
The live stream description notes we'll be hearing from Sam Altman, Gabriel Goh, Prafulla Dhariwal, Lu Liu, Allan Jabri, and Mengchao Zhong discussing 4o image generation.
(Image credit: Future)
OpenAI debuts native image generation in the 4o model
(Image credit: Future)
And we're off to the races – Sam Altman is calling this one of the most fun advancments, and it's native image generation in the 4o model. He quickly noted 'it's a huge step forward' and something that OpenAI has been excited to rollout for quite some time, for a whole host of folks.
Altman notes the best way to explain it is to show it off, so we're already in a demo. In just a few seconds after the prompt, OpenAI showed off an image with what the team said has 'perfect text.' Seemingly showing a leap in terms of understanding the prompt and creating the image with clear text, and a unique point of view effect.
(Image credit: Future)
(Image credit: Future)
In the second demo, the OpenAI team took a selfie and then asked ChatGPT to make it into an 'anime style.' It took several seconds, but it did indeed generate what was requested. You can see it above.
Sam Altman was then quick to note that the improved image generation is starting to roll out now in ChatGPT and Sora for Pro users, and it will be available for free users as well.
We also are seeing the process of the native image generation model within the 4o model, turning that generated selfie into an "AGI meme."
(Image credit: Future)
Sam Altman also teased that the native image generation model within 4o is designed to be a little offensive within reason if that's what you direct it to. The key phrase there is "within reason," and no doubt many users will put that to the test.
Now, the second demo asks for a colorful image describing the theory of relativity, with some added humor. Altman also noted that the image generation model is a bit slower but that the result is much higher in quality.
(Image credit: Future)
Considering the improved image generation is already available – or at least rolling out – TechRadar's editor-at-large, Lance Ulanoff, already tested the feature.
Lance took a selfie and uploaded it to ChatGPT via the iPhone app. He then asked for it to be turned into anime style. The first time, it gave him a full head of hair, but then corrected when he asked for it to be bald.
(Image credit: Future/Lance Ulanoff)
(Image credit: Future)
Back to the live demos, OpenAI is showing that we can now chat with ChatGPT more visually. This means that you can ask for requests to images in a row, and it will remember the context.
In this example, a photo of a coin was sent, and then the team asked ChatGPT to make it transparent, among other requests.
(Image credit: Future)
That was fast!
OpenAI certainly covered quite a bit of ground in just about 15~ minutes. Sam Altman and the team debuted native image generation in the 4o model. Then, presented some demos, and before it was wrapped, we already tested the feature in the ChatGPT app for the iPhone.
Now, as OpenAI announced, the improved model is rolling out now to Pro users, but is also coming to free users. Altman also confirmed it will eventually arrive in the API as well.
We just put image generation in the 4o model through another test, this time asking for a cartoon strip in the style of Charles Schulz's "Peanuts." While ChatGPT acknowledged the request, it turned it down due to copyright.
Instead, the resulting funny comic strip is in a similar style, with two familiar characters who have new names and other qualities to distinguish them from the original.
(Image credit: Future)
Now that OpenAI's improved image generation has been out for the public to use for several hours let's walk through how it's been so far.
On my side, I have a free ChatGPT account and quickly used up my daily allotment in the span of only a few minutes, generating a few images of a dog. When I asked for a fourth, I got the response:
"It seems like I can’t generate any more images right now. Please try again later. Let me know if there’s anything else I can do for you!"
As with most new features, the free tier of ChatGPT will have some limitations.
My colleague Lance Ulanoff, who's on ChatGPT Plus, has had better luck, though the time frame for generating images has stretched far, in some cases generating an alert that the request timed out and to retry, though the image was ultimately generated.
That could result from the network connection, potentially heavy loads of interest on OpenAI's servers, or even a limit ... though the latter doesn't seem as likely for a paid user.
