OpenAI announced GPT-6 Astra on September 3, 2026. The model is aimed at multi-step work across browsers, code, tools and professional software, rather than simply producing longer chat answers. OpenAI's documentation says it is rolling out first to enterprises in its Trusted Access Program, with the API and Plus, Pro, Business and Enterprise plans coming in the following days.
What changes with GPT-6 Astra
Astra is designed for tasks that must preserve context, use tools repeatedly and adjust the plan as results arrive. It supports computer use, web search, file search, code interpreter, hosted shell, MCP and image generation. In the Responses API, developers can also use asynchronous tool calls so the model can continue other work while an external tool is running.
Another practical change is mid-turn steering. An application can send new instructions while the model is working, allowing Astra to preserve completed work and continue with the correction. Developers can also change reasoning effort within the same conversation without resending the entire prompt. These features matter more for coding, document work, browser operations and long-running agents than for short questions.
Who can use it now?
The confirmed access path is an initial rollout to enterprises in the Trusted Access Program. OpenAI says the API and Plus, Pro, Business and Enterprise plans will follow in the coming days. The exact timing may vary by region, account and product surface, so a launch announcement does not mean every account can switch today.
On the API, set the model to gpt-6-astra. OpenAI recommends the Responses API for the full tool workflow. If an existing integration uses none or minimal reasoning, migration should start at low and be measured. Astra does not support none, and Fast mode is unavailable with EU data residency.
The price is high, so task-level efficiency matters
OpenAI lists API pricing per one million tokens at $10 for input, $1 for cached input, $12.50 for cache writes and $50 for output. The context window is 1,050,000 tokens and the maximum output is 128,000 tokens. By token price alone, Astra is not a cheap model.
OpenAI's migration guidance argues that Astra can complete some tasks with fewer output tokens, so teams should measure the cost of finishing a real job instead of comparing only per-token rates. For users, the practical question is whether it reduces errors and handoffs in browsing, coding and document work. Enterprises must also account for permissions, audit records, data residency and high token costs.
More capability also requires stronger governance
OpenAI describes Astra as its strongest model for computer use and professional work, alongside misalignment monitoring and stronger cyber safeguards. Its safety documentation also discusses a Critical cybersecurity capability designation. The direction is clear: more independent completion of complex work, but high-impact actions, external tools and security workflows still need scoped permissions, human review and auditability. Adoption should be based on a team's workflow and error costs, not launch-time benchmark claims alone.
OpenAI documentation: GPT-6 Astra model
