GitHub made GPT-6 Astra generally available in Copilot on 4 September 2026. This OpenAI model, described as built for long-running coding work and autonomous chains of steps, joins the model picker in the paid tiers. For web teams, the point is not one more model in a list, but an agent able to carry a refactor end to end before handing back control.
A model built for long tasks
According to GitHub’s changelog, GPT-6 Astra plans and validates as it goes, batches diagnosis with verification, then confirms its own result before declaring a task done. That internal loop targets the cases where an agent must hold a goal across several steps: migrating a module, fixing a cross-cutting regression, updating a test suite. It is the same direction already taken by editorial agents in production, moved over to the developer’s workstation.
Where the model shows up, and for whom
GPT-6 Astra is reserved for the Copilot Pro+, Max, Business and Enterprise tiers. It is selected from the same picker as the other models, across a broad range of surfaces.
| Copilot tier | Access to GPT-6 Astra |
|---|---|
| Free / Pro | Not included |
| Pro+ | Available in the picker |
| Max | Available in the picker |
| Business | Available, subject to the administrator’s model policy |
| Enterprise | Available, subject to the administrator’s model policy |
On the tooling side, the model is offered in Visual Studio Code, Visual Studio, JetBrains, Xcode, Eclipse, the command line via Copilot CLI, the Copilot coding agent, the GitHub app, github.com and GitHub Mobile on iOS and Android.
Model picker -> GPT-6 Astra
Eligible tiers : Pro+, Max, Business, Enterprise
Surfaces : VS Code, Visual Studio, JetBrains, Xcode, Eclipse,
Copilot CLI, coding agent, github.com, GitHub Mobile
Billing : usage-based, at the provider's list priceUsage-based billing and administrator control
GPT-6 Astra is billed by usage, at the provider’s list price, and is not part of a request allowance. On Business and Enterprise, access stays governed by the model policy defined in Copilot settings: an administrator can enable or block it for the organisation before teams ever see it appear. That guardrail matters, because an agentic model consumes far more than a plain completion, and its cost tracks the number of steps it chains together.
A model that confirms its own results is no substitute for reading what it produces.
Watch out. Usage-based billing makes the cost of an autonomous agent hard to predict: a poorly scoped task can multiply the steps, and so the bill. Setting clear boundaries, capping access through the model policy and reviewing every diff remain the only reliable safeguards.
The bottom line
GPT-6 Astra is not a faster completion but an agent-style model aimed at long tasks, open only to the paid Copilot tiers and billed by usage. Its arrival confirms that the developer’s workstation is shifting from the copilot that suggests to the agent that executes, with all the cost and review questions that raises. The good practice fits in one line: scope before launching, review before merging.
I have been testing these agent models on real code for months, and my take hasn’t changed: they save time on well-marked work, and they lose it the moment you hand them an architecture decision without a guardrail. GPT-6 Astra validating its own results is useful, but I never sign off on a diff I haven’t read. The day I did, it would no longer be my code. — Simon Janvier
Further reading: the official changelog, GPT-6 Astra is generally available in GitHub Copilot (GitHub Changelog).
