OpenAI launched ChatGPT Images 2.0 yesterday — and the model behind it is called gpt-image-2. The headline feature: the system no longer jumps straight into image generation. Instead, it works through the prompt in stages first. OpenAI calls this “Thinking.”
What ‘Thinking’ Means for Images
Essentially, OpenAI applied its reasoning approach to image generation. When you send a complex prompt — say, a diagram with labels in three languages — the model first analyzes which elements are needed, pulls in real-time information if necessary, and then generates the image. Sounds like a small feature, but it meaningfully changes result quality.
The Hard Facts
Resolution goes up to 2K (2,576 pixels on the long edge). Text in images gets dramatically better — OpenAI demonstrates correct rendering of Bengali and Hindi, which has been a real weakness until now. Multiple images within a session can maintain consistency, which is enormously helpful for brand assets or storyboards.
All ChatGPT and Codex users get access, with paid users able to generate more advanced outputs. For developers, the model is available via the API as gpt-image-2.
How This Fits Into the Bigger Picture
OpenAI has always been strong in image generation, but lost ground to Midjourney and Google’s Imagen in recent months. Images 2.0 feels like a targeted response: rather than just improving image quality, OpenAI changed the entire pipeline approach.
Text rendering in particular is a real differentiator. Anyone who’s ever tried to generate a poster with readable text using DALL-E knows how frustrating that’s been. If gpt-image-2 truly solves this reliably, it’s a major step forward.
For the Claude world, this matters because Anthropic doesn’t offer its own image generation model. Claude Design creates visual prototypes and designs, but it’s based on code output (HTML/SVG), not an image model. OpenAI is playing in a league where Anthropic isn’t competing — yet.
Sources: