Introducing GPT-6 Astra for developers
- Source
- simonwillison.net
- Date
OpenAI rolled out GPT-6 Astra for developers this week, and the headline changes are prompt comprehension and 3D rendering. Simon Willison ran his standard pelican-on-a-bicycle test against it and found the output noticeably sharper than prior releases, with better handling of proportion and pose. The bigger shift is in 3D-model and complex-scene generation, where Astra produces structured geometry and lighting that earlier image models routinely got wrong. For teams building agents that call an image or video generation API, that means fewer manual corrections when a scene needs consistent depth, occlusion, or multiple interacting objects. It also raises the bar for anyone benchmarking image generation quality, since Astra's 3D fidelity gives reviewers a harder target to test against. Developers evaluating vendors for 3D or complex-scene work should factor this release in before committing to an existing provider.
Context
Long-running AI tasks need a way to accept changes without starting over. In its Astra developer video, OpenAI describes combining asynchronous tool calls with steering in the Responses API. The model can continue other work while an outside operation runs, then use its result when available. Steering lets the user add or change direction during that response.
In the publisher's described example, a new instruction changes the model's next steps while an already-started tool keeps running. The distinction is between redirecting subsequent work and cancelling an operation already underway. This explains the announced interaction design; the promotional example does not establish its reliability across tasks.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…