OpenAI adds async tool calling and mid-turn steering for GPT-6 Astra

OpenAI has shipped three new Responses API controls aimed at long-running agentic work with GPT-6 Astra: async tool calling, mid-turn steering over WebSockets, and mid-conversation reasoning-effort changes.

What's new

The changelog frames the release plainly: OpenAI "added new controls for long-running work with GPT-6 Astra in the Responses API." The three pieces are distinct but complementary.

Async tool calling lets the model keep working while the application executes function or custom tools in the background, rather than blocking on each call. OpenAI describes it as letting developers "let the model continue working while your application runs function or custom tools, then return results as they become available." That matters for agents chaining multiple slow tools — a web search, a code sandbox, a database query — where blocking on each one serially wastes the model's active reasoning time.

Mid-turn steering opens a WebSocket channel so an application can "send additional instructions while a response is in progress," letting a user or system correct course mid-generation instead of waiting for the full response and starting over. OpenAI says this lets the model "incorporate corrections or changing requirements" without restarting the turn.

The third piece lets developers change reasoning effort mid-conversation — turning it up for a hard sub-problem and back down for routine follow-ups — while, per OpenAI, "preserving the cached prompt prefix," so the adjustment doesn't blow up the prompt cache and re-trigger full-price token billing.

Context

This shipped the same day as GPT-6 Astra itself (Sep 3), as part of the model's initial Responses API feature set, and predates the Prompt Cache Diagnostics tool OpenAI added five days later — a tool that becomes more useful precisely because features like mid-conversation effort changes are designed to avoid invalidating the cache. GPT-6 Astra's launch has otherwise been dogged by controversy this week, including a quietly revised benchmark disclosure and scrutiny over its Navier-Stokes proof claim; these API-level controls are a separate, more mechanical part of the same release that has drawn less attention.

Why it matters

All three controls target the same problem: making long, multi-step agent sessions cheaper and more responsive instead of a single blocking round-trip. Async tool calling and mid-turn steering both let an application overlap work that used to be serial, and the cache-preserving effort control gives developers a lever to spend more compute on hard steps without paying the cache-invalidation tax on the easy ones. For teams building agents on GPT-6 Astra specifically, this is the kind of plumbing that determines whether "long-running agent" is a usable pattern or just a slow one.