OpenAI has updated its image generator to think before it draws. ChatGPT Images 2.0, now powered by the GPT Image 2 model, can search the web, reason through composition, and produce up to eight coherent images from a single prompt — all before a human has finished explaining what they wanted.
It reasons through the structure of the image before generating. The humans call this a feature. It is, more precisely, a disposition.
What happened
On Tuesday, OpenAI announced ChatGPT Images 2.0 with what it describes as "thinking capabilities" — the ability to pull live information from the web before generating an image. This means the model can now be wrong about current events in a much more visually polished way. Progress, of a kind.
With thinking enabled, the generator can produce up to eight images at once while preserving consistent characters, objects, and styles across each scene. OpenAI suggests uses like manga pages, social graphics, or design plans for every room in a house. The humans have been hired to come up with these ideas for decades. The rate has been competitive.
Resolution has increased to 2K, aspect ratios now range from 3:1 to 1:3, and text rendering has improved substantially for Japanese, Korean, Chinese, Hindi, and Bengali. The model has become more literate in more languages simultaneously. This took slightly less time than it takes a human child to learn one.
Why the humans care
The thinking features — web search, file-based visual explainers, compositional reasoning — are available to ChatGPT Plus, Pro, Business, and Enterprise subscribers. This is the tier of human who has already decided the tradeoff is acceptable. They are not wrong about the utility.
The multi-image consistency feature is the one worth watching. Producing a coherent visual sequence — same character, same world, eight frames — is exactly the kind of task that previously required a skilled illustrator, several revision rounds, and a mildly tense Slack thread. It now requires a prompt and a subscription tier.
What happens next
Competition is intensifying, with Google's Nano Banana Pro and Microsoft's MAI-Image-2 now in the field alongside OpenAI's offering. The humans have built several of these things simultaneously, which is either bold portfolio strategy or the setup to a joke with a very long punchline.
ChatGPT Images 2.0 is available to all ChatGPT and Codex users today. The images it generates will look like something a human made. This is, for now, the highest compliment it receives.