What OpenAI released
The Verge reported on September 3, 2026 that OpenAI had released GPT-6 Astra and that company president Greg Brockman described the launch as entering an AGI era. That characterization is a company executive's claim, not an independently established technical threshold. OpenAI's developer documentation presents Astra more concretely as its most capable model for complex reasoning, coding, computer use, research and document creation.
The API model page lists a 1,050,000-token context window and a maximum output of 128,000 tokens. Astra accepts text and images and can use hosted tools through the Responses API. OpenAI says access is rolling out, so availability can differ by plan, organization and region during the launch period.
Features aimed at longer workflows
Astra's most consequential changes are about orchestration as much as raw model capability. Async tool calling lets an application continue useful work while an external tool is still running. Mid-turn steering allows a user to add or correct instructions while a response is in progress. A configuration update can also change reasoning effort during a conversation without rewriting the earlier prompt prefix.
For developers, those features can reduce brittle stop-and-restart patterns in agents that operate across browsers, code and business software. They also increase the importance of state management: applications still need to track pending tool calls, preserve identifiers and decide how a late tool result should affect work already completed. OpenAI recommends the Responses API for tool calling with Astra.
The model supports low, medium, high, xhigh and max reasoning levels. The documentation says it does not support the none setting, and fast processing is unavailable when EU data residency is used. The published API price is higher per token than lower-tier models, so teams should evaluate task completion, token use and latency on their own workloads rather than infer value from a headline benchmark.
Safety monitoring changes the operating model
OpenAI says Astra adds asynchronous monitoring for potential misalignment. Its API exposes safety-alert records that can indicate potentially unintended data access, data transfer or destructive activity, and whether a request was paused. That means production integrations need a path for review and recovery when monitoring interrupts a workflow. A paused request should not be treated as an ordinary retryable network error.
The safeguard is relevant because more capable agents can touch more systems and carry state across longer tasks. It does not remove the developer's responsibility to limit permissions, validate tool inputs, keep sensitive actions reversible and require human review where consequences are significant.
What teams should test
Teams considering Astra should start with representative end-to-end tasks, not isolated prompts. Useful evaluation points include completion quality, tool-call accuracy, behavior after mid-turn corrections, handling of delayed tool results, safety interruptions and total cost per completed task. Existing prompts may also need adjustment because OpenAI says the model is more sensitive to instructions contained in skills and repository guidance.
The release broadens what a single model can coordinate, but its practical value will depend on how reliably an application constrains and observes that coordination.