GPT-6 Astra API Update: Async Tool Calling and Mid-Turn Steering

GPT-6 Astra shipped with API changes that matter more to agent builders than the model itself. Async tool calling, mid-turn steering, mid-conversation effort changes and cross-tool computer use all change how you structure an agent loop. This post covers what each does and what to change in an existing agent.

Sources: OpenAI's model page (developers.openai.com), OpenRouter's model listing, and OpenAI's own API update email to our inbox dated Sept 18, 2026. We have not benchmarked any of this ourselves.

What did the September 2026 GPT-6 Astra API update change?

Four things changed for agent builders. Tools can run asynchronously while the model keeps reasoning. You can send corrections while a turn is running. You can change reasoning effort mid-conversation without breaking the cached prompt prefix. And computer use can drive tools that have no API. The first and last two come from OpenAI's email; the docs list mid-turn steering and WebSocket mode as their own guides.

The model itself has a 1,050,000 token context window and 128,000 max output tokens. It accepts text and image input and returns text. Reasoning effort supports low, medium, high, xhigh and max (OpenAI model page, checked 2026-09-20).

How does async tool calling work?

Setting async: true on a function or custom tool lets the model keep reasoning while your app runs the tool. Before, a slow tool call stalled the turn. Now the model can continue thinking about other parts of the task and pick up the result when it arrives. This detail is from OpenAI's email. We have not seen the full schema, so check the docs before you rely on exact field behavior.

Where this helps: agents that call a slow search, a database export or a browser step while there is other work to plan. Where it does not: tools whose output the next step depends on. There, async gains you nothing.

What is mid-turn steering?

Mid-turn steering lets you send a correction to the model while it is still working, instead of waiting for the turn to end or cancelling it. The email says this is reportedly done over WebSocket. OpenAI's docs list "Mid-turn steering" and "WebSocket mode" as separate guides under the Responses API, but we did not read their contents in detail.

The practical value is cost. A wrong direction caught at minute one no longer burns a full turn of output tokens at $50 per 1M.

Can you change reasoning effort mid-conversation?

Yes, and OpenAI says the prompt prefix stays cacheable when you do. That lets you run early planning steps at high effort and routine follow-ups at low effort in one conversation, without paying for a cold cache each time. This is from the email; we could not find the caching rule spelled out on the model page, so test your cache hit rate before assuming it.

What does GPT-6 Astra cost?

Standard rates are $10 per 1M input tokens, $1 cached input, $12.50 cache writes and $50 output. Prompts above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output for the full request.

Cached input costs a tenth of fresh input, so cache-friendly loops matter. Output is five times input. Long, chatty agents cost more than long prompts.

OpenRouter lists the release as Sep 4, 2026. Other reports say Sept 3. We could not confirm which.

What should you change in an existing agent?

Who is this not for?

If your agent makes short, single-shot calls with fast tools, none of this changes much. Async and steering pay off in long, multi-step runs. Teams without a WebSocket-capable client should also read the steering docs first, since the setup differs from plain request and response.

How does this compare with Claude?

For model routing between Astra and Claude, see our draft comparison once it is live. For coding agents specifically, our Codex vs Claude Code post covers the trade-offs, and what is Codex CLI and how to use Codex cover the OpenAI side.

FAQ

Does async tool calling work with every tool?

The email says async: true applies to function and custom tools. We have not confirmed other tool types.

Is mid-turn steering the same as cancelling a run?

No. It sends a correction while the model keeps working. Cancelling ends the turn.

Does changing reasoning effort break prompt caching?

OpenAI's email says the prefix stays cacheable. Verify with your own cache-hit metrics.

When do the higher long-context rates apply?

Above 272,000 input tokens, the full request bills at 2x input and 1.5x output (OpenAI model page, 2026-09-20).

Next step

Pick your slowest agent, tag one independent tool async and measure the turn time. If you want help redesigning a production agent around these features, see our AI automation agency or book a consultation.