The important OpenAI change is not just a new model – it is a more steerable agent runtime
OpenAI introduced GPT-6 Astra on September 3, 2026. The headline is higher model capability, but the more important operational shift is what Astra adds to long-running agent workflows.
Astra is being rolled out to ChatGPT Plus, Pro, Business, Enterprise, and the API. OpenAI positions it for complex end-to-end work across coding, research, computer use, documents, spreadsheets, and presentations. API pricing is $10 per million input tokens and $50 per million output tokens, so this is not a model to route every task through automatically. The practical question is whether fewer failures, fewer human interventions, and fewer reruns make the cost per finished result lower.
Sources:
The bigger architecture change: async tools and mid-turn steering
The Responses API adds three Astra capabilities that can change how agent systems are designed.
- Async tool calling: the model can continue useful work while external tools run.
- Mid-turn steering: a user can send a correction or new requirement while the response is already in progress over WebSockets, without discarding completed work.
- Dynamic reasoning effort: reasoning can be raised for a hard decision or lowered for routine follow-ups while preserving the cached prompt prefix.
This moves agent design away from a rigid sequence of model call, tool call, wait, model call. A system can become partially concurrent, interruptible, and more economical about where it spends reasoning effort.
There are constraints. Astra tool calling requires the Responses API. It does not support none reasoning effort or custom temperature and top_p controls. Fast mode is unavailable for Astra with EU data residency. OpenAI also adds asynchronous misalignment monitoring that can alert or stop supported agent work for review.
Source:
Enterprise computer use gets a real policy layer
OpenAI also added more granular enterprise controls for Codex browser and computer use. Admins can define website rules, restrict uploads and downloads, control approval persistence, and allow or block native macOS and Windows applications. User approvals cannot override an administrator restriction.
That matters because serious desktop agents need more than a good prompt. They need enforceable boundaries. For enterprise deployments, agent governance is moving into the platform control plane.
Sources:
- https://openai.com/products/release-notes/
- https://help.openai.com/en/articles/20001510-manage-browser-and-computer-use-in-your-enterprise-workspace
ChatGPT Sites gets a private external-sharing layer
Eligible ChatGPT Site owners can now invite named external viewers without making the Site public or adding those people to the workspace. This creates a useful middle ground for client portals, stakeholder dashboards, and review prototypes.
The catch is regional: OpenAI’s current documentation says ChatGPT Sites is not available in the EEA, Switzerland, or the United Kingdom at launch. That makes the feature interesting but unavailable for many European users today.
Source:
What to test next
The useful Astra benchmark is not “is it smarter?” It is:
How much does it cost to get one accepted end-to-end result after counting interventions, corrections, retries, and tool latency?
That metric will determine whether Astra belongs at the top of a model router or only in a small set of high-value workflows.
