OpenAI ships GPT-6 Astra to the API at $10/$50 per million tokens
OpenAI's new flagship reaches the API and Bedrock with a 1M-token window, computer use, and live mid-turn steering — cyber exploit work stays gated.
Copy markdown
The price, and where you can call it
gpt-6-astra runs $10 per million input tokens and $50 per million output ($1 cached input), with a 1.05M-token context window and 128K max output. It's rolling out now on the OpenAI API and Amazon Bedrock; ChatGPT Plus, Pro, Business, and Enterprise get it within days.
Mid-turn steering: redirect a run without restarting it
Over a WebSocket to the Responses API you can now inject new instructions into a response while it's still generating — call response.create, then response.steer with the response ID. The model finishes its current step and continues with your update; it won't undo output already sent or cancel tools already running. Astra-only for now.
Async tools and mid-run reasoning control
The Responses API also adds async tool calling, so function calls run concurrently instead of blocking the model, plus the ability to change reasoning effort mid-conversation while keeping cached prefixes intact. reasoning.effort now scales from low to max.
The coding jump, in numbers
Astra posts 57.7% on Terminal-Bench 4.0 (up from 37.3% on GPT-5.6 Sol) and 72.6% on OSWorld 2.0, finishing those computer-use tasks about 47% faster. Tool support ships day one: web search, file search, code interpreter, computer use, MCP, and skills.
Watch the long-context bill
Prompts over 272K input tokens flip to 2x input/cache and 1.5x output rates, and a Fast mode doubles the price for up to 2.5x speed. Knowledge cutoff is April 30, 2026, and enterprise admins must switch Astra on — it's off by default.