GPT-6 Astra Hit Copilot in a Day: Test It on Your Repo First

OpenAI introduced GPT-6 Astra on September 3, and one day later GitHub made it generally available in Copilot for Pro+, Max, Business and Enterprise plans. A frontier model can now appear in your developers' model picker the day after the announcement, before anyone on the team has decided whether it belongs there. That short gap between launch and adoption matters more to me than the benchmark chart.

What actually shipped#

• The API model is gpt-6-astra. Standard pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. Fast mode in the API offers up to 2x the speed at 2x the Standard price.
• It starts with a limited set of organizations, then rolls out over the following days to ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API, Microsoft Azure and AWS Bedrock.
• The coding headline is Terminal-Bench 4.0: 57.9%, against 55.8% for Claude Fable 5.1 and 37.3% for GPT-5.6 Sol. OpenAI says it gets there at roughly 9% lower estimated API cost per task than Sol and 63% lower than Fable 5.1.
• In Codex, Astra can keep notes across context windows instead of relying only on compaction, and earlier windows stay searchable. It is experimental and opt-in through config.toml, and OpenAI says it becomes the default for Astra in the coming weeks.
• In Copilot it covers VS Code, Visual Studio, the Copilot CLI, the coding agent, JetBrains IDEs, Xcode, Eclipse, github.com and more, billed at provider list pricing under usage-based billing.

The defaults differ by surface#

This is the part to check on Monday. On the OpenAI side, Enterprise administrators can enable Astra for their workspace, and access is off by default at launch. On the GitHub side, the changelog describes default model enablement: new models are turned on automatically unless an administrator has switched off the global default or explicitly disabled this model.

So if your organization is on Copilot Business or Enterprise and nobody has touched the model policy, Astra can show up in the picker as the gradual rollout reaches you, billed at list price. Neither default is wrong, but know which one applies to you before the usage report tells you. Confirm whether the global default is on and that spending alerts exist for usage-based billing.

Why the vendor chart is not your decision#

The benchmarks are real signal, and OpenAI's fine print is worth reading. Its scores are the maximum at any effort level, and its GPT evaluations ran in a research environment or through the API, which may differ slightly from production ChatGPT. Your harness differs again: conventions, CI, flaky tests, instructions file, review bar. A 2.1-point lead over Fable 5.1 measured in someone else's harness is not enough for me to change a team default.

How I would evaluate it on a real repo#

• Take 15 to 20 recently merged PRs of different kinds: a bug fix, a feature, a refactor, a migration, a missing test. Reset to the parent commit and give the model only the issue text.
• Keep the harness, instructions and effort level the same for every candidate. Change one variable at a time.
• Score what you care about: tests pass, conventions followed, diff size versus the human diff, review rounds, tokens and cost per task, wall-clock time.
• Put every diff through the same review gate. I use Quorate for this, my CLI that runs several AI reviewers over a diff and applies deterministic PASS/WARN/FAIL gates, so the comparison does not depend on my mood that day.
• Include at least one long task where an early constraint matters late, like a multi-step refactor. That is where the notes mode should beat plain compaction. Per the config reference it is off by default and needs ChatGPT sign-in on Plus, Pro or Pro Lite, so API-key setups miss it:

[features.context_management]
experimental_mode = true

• Record where it stops. OpenAI says extra safety checks can pause or stop legitimate work. In ChatGPT or Codex you may be asked to review the action, but in the API the task simply stops. In unattended pipelines, treat that as a normal failure mode, not an edge case.

There is also a security boundary. OpenAI says Astra meets the Critical cybersecurity threshold under its Preparedness Framework. The launch version supports tasks like secure code review and patching but refuses more advanced ones, such as creating proof-of-concept exploits. OpenAI plans to widen access through Daybreak in the coming weeks. If your security team does exploit validation, this is not the tool for it yet.

What I'm doing about it#

I already route work across several coding agents by task type and cost, so a launch does not change my default. It adds a candidate. This week I will pick a set of merged PRs from my own projects, run Astra next to my current defaults, and compare cost per accepted diff, not cost per token. At $50 per million output tokens, a verbose agent loop gets expensive quickly, so accepted changes per dollar is the number to watch.

For teams, my advice is short: check your Copilot model policy today, decide who should have access, and only change the default once the model has won on your own repo.