OpenAI ships GPT-6 Astra: 1.05M context, $10/$50, async tools
OpenAI's most capable model brings async tool calling and mid-turn steering; tool use needs the Responses API, and access is gated to top tiers first.
Copy markdown
The model, and the bill
GPT-6 Astra is OpenAI's new flagship: a 1,050,000-token context window, 128K max output, an April 30, 2026 knowledge cutoff, and reasoning effort dialable from low to max. It runs $10 per 1M input and $50 per 1M output (cached input $1) — roughly 2.5x the outgoing promo pricing.
Async tool calls, mid-turn steering
The Responses API gained three things aimed at long agent runs: async tool calling (the model keeps working while your app resolves a function call), mid-turn steering (push new instructions over WebSockets during a response), and changing reasoning effort mid-conversation without dropping your cached prompt prefix.
Tool calling needs the Responses API
Astra is exposed on Chat Completions, Responses, Realtime and Batch, but tool calling runs only through the Responses API — and it rejects `none` reasoning effort, custom temperature/top_p, and logprobs. Move agent code off Chat Completions if you want tools.
Where it wins, where it doesn't
At max effort Astra posts 74.1% on DeepSWE v1.1 coding — but Meta's Muse Spark 1.3 edges it at 75.4% — alongside 97.6% on FrontierMath Tier 4 and 64.6% on Terminal-Bench Science vs Fable 5.1's 52.6%. Real gains on science and agentic tasks, not a clean sweep on code.
You can't call it yet unless you're gated in
Rollout is enterprise and Trusted Access first, with Plus, Pro, Business and broader API access promised 'in coming days.' Fine-tuning isn't supported. Plan for staged availability, not day-one access.
The safety asterisk
OpenAI says Astra's reasoning was harder to monitor than the prior Sol generation in adversarial tests, ships asynchronous misalignment monitoring on API calls, and will withhold further scaling until its monitoring confidence improves.