OpenAI releases GPT-6 Astra for end-to-end agent work

OpenAI’s September 3 developer changelog introduces GPT-6 Astra, a model for end-to-end work across reasoning, coding, computer use, research, and document creation. The release is paired with new Responses API controls for applications that need to keep work moving while tools run.

A model for multi-step work

OpenAI describes GPT-6 Astra as a model that can carry a complex task from an initial request to a finished result using the context and tools provided by the application. It is available through v1/responses and v1/chat/completions.

The model has several migration constraints: it does not support the none reasoning-effort level, custom temperature or top_p, or logprobs. Tool calling requires the Responses API, so applications using tools through Chat Completions should follow OpenAI’s Responses migration guidance rather than changing only the model name.

Controls for long-running Responses API work

The same changelog entry adds asynchronous tool calling, letting the model continue while an application runs a function or custom tool and returns results later. It also adds mid-turn steering over WebSockets, so an application can send additional instructions while a response is in progress.

Applications can change reasoning effort during a conversation while preserving the cached prompt prefix. That gives an orchestration layer a way to spend more compute on a difficult step and reduce effort for routine follow-ups without restarting the entire context.

Safety and operational details

OpenAI says supported Responses API requests can use asynchronous misalignment monitoring during agent work. Checks can trigger safety alerts or stop a conversation for review. Teams should account for those possible interruptions in state management, retries, and user-facing progress reporting.

The API changelog also documents distinct 429 slow_down and 503 server_is_overloaded errors for traffic increases and temporary model overload. When Retry-After is present, applications should wait at least that long; otherwise, use exponential backoff rather than treating both responses as the same failure.

Migration checklist

Before switching a production workflow, verify the endpoint, reasoning-effort values, sampling parameters, tool-calling path, and streaming behavior. Add tests for asynchronous tool completion, mid-turn corrections, safety-review stops, and the new rate-limit error codes. Keep the official model and API changelog in the runbook because the release is new and model behavior can evolve.

Sources

• OpenAI Developers — “Developer changelog” (September 3, 2026): https://developers.openai.com/api/docs/changelog
• OpenAI Developers on X — official developer updates: https://x.com/OpenAIDevs