What's new
Three notable improvements over the previous generation:
- **Text in images** — legible, clean text, no garbled gibberish. Text rendering has been a chronic weakness for all generative models; this is a noticeable step forward.
- **UI generation** — describe an interface in plain language and get a working mockup. The recommendation feed on the right came out cleanly.
- **Live web data** — the model pulls current information during generation, not just what's in its training data.
There's also extended thinking mode: give it a short prompt and the model comes up with a concept on its own and delivers a finished image — no detailed instructions needed.
Non-standard formats
OpenAI also showed how the model handles arbitrary layouts — ad banners, multi-column spreads, full newspaper pages with headlines and body copy:
Why it matters
Accurate text in images is what kept designers from using generative models for real work: logos, banners, UI mockups. Now it works. All in all — a genuinely solid release from Altman and team, nothing to be embarrassed about.
Expecting GPT-5.5 or a new model called Spud within the next couple of weeks.
OpenAI — try it via "try in ChatGPT"; works in the browser version, not in the app yet.
TechCrunch