OpenAI bakes image gen into GPT-4o, and it can actually read

OpenAI just made image generation a native feature of GPT-4o, and the early details suggest it’s a genuine step beyond the dreamlike chaos of earlier models. The company has long argued that visuals should be a core language model capability, not an afterthought, and this release is the proof. The headline upgrade is text rendering that actually works, turning the model into a tool for logos, diagrams, and infographics rather than just surreal art. It’s a shift from decoration to communication.

Under the hood, the model was trained on the joint distribution of online images and text, learning not just what a cat looks like but how images relate to each other and to language. Combined with what OpenAI calls ‘aggressive post-training,’ the result is a system that can handle prompts with 10 to 20 distinct objects while binding their traits and relationships correctly. Other systems, the company notes, typically start to fall apart after 5 to 8 objects. The model can also analyze uploaded images and use them as visual inspiration, maintaining consistency across multiple iterations in a chat—a character’s face stays the same as you tweak their armor, for instance.

On the safety front, OpenAI is deploying a reasoning LLM trained on human-written safety specs to moderate both text prompts and output images, an approach similar to its deliberative alignment work. All generated images carry C2PA metadata for provenance, and the company has built an internal search tool to verify whether content came from its model. Heightened restrictions apply to images of real people, with particularly robust safeguards around nudity and graphic violence. The team acknowledges the model isn’t perfect and plans to iterate on known limitations post-launch.

The feature rolls out starting today as the default image generator for Plus, Pro, Team, and Free users. That wide availability signals confidence, but the real test will be whether 4o’s visual fluency holds up when millions of users start stress-testing its ability to follow precise, multi-object instructions in the wild.