GitHub Copilot Review (September 2026): GPT-6 Astra, Four Models Retiring, and the Usage Billing Math

GitHub Copilot shipped nine changes in the first four days of September, and the biggest one quietly changes what a seat costs. GPT-6 Astra went generally available on September 4, OpenAI's long-horizon autonomous coding model, and it bills at provider list pricing under usage-based billing rather than riding for free on your monthly seat (GitHub changelog, Sept 4). Our August 24 review rated Copilot 4.5 and predated all of it. The re-call after this week: 4.5 out of 5, held, and the reasoning below is specific.

Here is the one thing to leave with. The frontier models Copilot added this month do not ride on the flat seat price. GPT-6 Astra and Claude Fable 5.1 both meter against your credit allowance at provider rates, so on a $19-per-user Business plan a team that points everyone at GPT-6 for long agent runs will burn the pooled allowance and then either pause or pay for the overage (GitHub changelog, Sept 4). The seat still bounds your headcount. It no longer bounds your model spend. Route deliberately and the September wave is a clear upgrade. Route lazily and it is a surprise line item.

What changed since August 24

Nine items landed between September 1 and September 4, each dated and with its plan gate. Note the release status on each, because two of the biggest are still gradual rollouts and one headline feature is public preview.

• GPT-6 Astra went GA (Sept 4). OpenAI's latest general-purpose model, aimed at long-horizon, autonomous coding. GitHub's own writeup says it plans and validates as it goes, batches diagnosis with verification, and confirms its own results before calling a task done, which it credits for stronger long-horizon performance in fewer steps than prior OpenAI models (GitHub changelog, Sept 4). Available on Pro+, Max, Business, and Enterprise. Rollout is gradual.
• Claude Fable 5.1 went GA (Sept 1). Anthropic's Mythos-class model for long-running coding and knowledge work, on Pro+, Max, Business, and Enterprise. It bills usage-based, and it carries a policy wrinkle covered below (GitHub changelog, Sept 1).
• Gemini 3.8 Flash went GA (Sept 3). The cheap-fast slot, and the only one of the three new models that reaches Pro at $10, not just the premium tiers. It bills at introductory provider pricing through December 31, 2026 (GitHub changelog, Sept 3).
• Four models were flagged for deprecation (Sept 3). Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 all retire October 2. The replacement table is in the next section (GitHub changelog, Sept 3).
• Copilot code review can approve PRs (Sept 1). Review now issues an approval assessment and, when an admin enables it, can submit an approval that counts toward the required-approvals rule (GitHub changelog, Sept 1).
• Business and Enterprise signups reopened (Sept 3). GitHub is gradually reopening sign-ups for card and PayPal customers after a gated period (GitHub changelog, Sept 3).
• Content exclusions reached the app and CLI (Sept 2). The Copilot app and CLI now respect content exclusion policies, so excluded files stay out of context across agentic surfaces, not just the IDE. Business and Enterprise only (GitHub changelog, Sept 2).
• Enterprise can set any model as the default (Sept 2). Enterprise-managed settings now pin any supported model as the org default for new conversations, and the default can vary by enterprise team (GitHub changelog, Sept 2).
• User budgets can expire (Sept 1). Admins can set an expiration date on an individual user budget, after which the user falls back to their cost-center or universal budget, cutting the manual cleanup of temporary overrides (GitHub changelog, Sept 1).

The usage-billing math is the decision this month

Every Copilot plan includes a monthly credit allowance, priced at 1 AI credit for $0.01, and all model usage draws from it (Copilot plans page, pulled 2026-09-07). That has been true for a while. What changed is that the two headline models this month, GPT-6 Astra and Claude Fable 5.1, are frontier-priced and both bill at provider list pricing under usage-based billing (Sept 4 changelog, Sept 1 changelog). A long agent run on GPT-6 does not cost the same as a chat turn on a lightweight model, and the difference lands on the same allowance.

For a solo developer on Pro+, this is manageable: you have a $70 monthly credit pool, and if you steer routine work to Gemini 3.8 Flash and reserve GPT-6 for the genuinely long-horizon tasks, the pool stretches (Copilot plans page, pulled 2026-09-07). Pro at $10 does not see GPT-6 Astra or Fable 5.1 at all, so the frontier decision starts at Pro+.

For a team on Business at $19 per user, the math is where the surprises live. Credits pool across the team, and admins decide whether paid usage continues once the allowance runs dry; if it does not, Copilot pauses until the next cycle (Copilot plans page, pulled 2026-09-07). Point ten engineers at GPT-6 Astra for all-day agent work and the pool empties, then either the work stops or the overage bills. The fix is not to avoid GPT-6. It is to treat it as a metered resource: set it as the reach-for model for long autonomous tasks, keep the everyday default on something cheap, and give admins a budget that reflects the real burn rather than the sticker price.

The model picker: what to default by task

The picker is now a routing decision, not a preference. Here is the pick by task rather than by benchmark, with each model's source.

One caveat on Fable 5.1 before you pin it as a default. Unlike other Claude models in Copilot, it requires data retention on by default so Anthropic can run its safety classifiers, and GitHub is explicit that retained prompts and outputs are not used to train Anthropic's models (Sept 1 changelog). Zero-retention access exists only for eligible enterprise customers. The policy also ships off by default, so a Business or Enterprise admin has to enable it before anyone can select the model. If your compliance posture assumes zero retention across the board, Fable 5.1 is an opt-in you make on purpose, not a default you inherit.

The October 2 deprecations: run this migration first

Four models leave every Copilot surface on October 2, including chat, inline edits, ask and agent modes, and completions (GitHub changelog, Sept 3). If a team has pinned any of them in a workflow, an integration, or an org default, the swap needs to happen before the date or the model silently stops serving. GitHub notes that Business and Enterprise admins may need to enable the replacement through model policy, since a replacement is not on automatically.

The admin checklist is short. Confirm which of the four your team actually uses, enable each replacement in the org or enterprise model policy, and verify the policy shows the new model in the picker before October 2. Four retirements in one shot is real churn for a team with pinned tooling, and it is the friction that keeps this month from being a straight upgrade.

Copilot code review can now approve pull requests

Before September, Copilot's code review commented. It could flag issues, but a human still had to click approve. As of September 1, Copilot adds an approval assessment to the overview comment on every review, and when an admin turns approvals on, Copilot can submit an approval that counts toward the repository's required-approvals rule (GitHub changelog, Sept 1). That closes a loop for teams that use Copilot as a first-pass reviewer.

Two guardrails keep this from being reckless. The feature is off by default, and control lives at three levels: enterprises can leave it off or defer to orgs, orgs can enable it broadly or per repository, and repository admins can even choose which file paths Copilot is allowed to approve (Sept 1 changelog). If new commits land after Copilot approves, its approval is dismissed the same way a human reviewer's would be. The obvious play for most teams: enable approvals on low-risk paths like docs and test fixtures, keep human sign-off on core logic, and let the assessment surface everywhere without letting it merge everywhere.

Pricing, still flat

Nothing on the sticker moved since August, which matters because the value went up while the price did not (Copilot plans page, pulled 2026-09-07). The variable now is the credit meter, not the seat.

All prices and allowances pulled from the Copilot plans page on 2026-09-07. New paid seats still enable gradually, and so do GPT-6 Astra and the other new models, so a fresh seat may not show the full picker on day one.

Rating

Copilot holds at 4.5 out of 5 in September 2026, unchanged from August. The hold is deliberate, not inertia.

The case to move up rests on the PR-approval capability, which is a genuine loop-closer for review workflows, and on three frontier models landing in one week (Sept 1 approval changelog, Sept 4 GPT-6 changelog). But approvals ship as public preview, off by default, so the value is potential until an admin wires it up (Sept 1). The case to move down rests on the usage-billing surprise for flat-rate Business teams and four deprecations in one shot (Sept 4, Sept 3). Neither of those is new in kind: Copilot has metered credits for months, and models rotate. The gains and the frictions land within a tenth of each other, so the score holds where August left it.

What still keeps it below a higher mark is the editing gap. Agent mode trails Cursor Composer on whole-repo multi-file refactors, the same edge that has held through prior reviews, and the credit model keeps monthly cost variable in a way a flat plan never was.

The pick, by who you are

Three buyers, three answers. September sharpens all three around the same lever: which model you route where.

Solo developer. Pro+ at $39 is the working tier, and GPT-6 Astra plus Claude Fable 5.1 are the reason to be on it rather than Pro (Copilot plans page, 2026-09-07). Wire Gemini 3.8 Flash to routine turns, reserve GPT-6 for the long autonomous runs, and the $70 pool holds. The flip condition has not changed: if your day is all-day whole-repo editing and you would rather not watch a meter, Cursor's Composer still edits cleaner on multi-file refactors and its flat usage removes the burn question. Our Cursor review covers those tiers.

Small team on Business ($19/user). The pick holds, with a budgeting caveat you should act on this week. Treat GPT-6 Astra and Fable 5.1 as paid usage on top of the pooled allowance, not as seat-included models, and set the team budget to match the real burn (Sept 4 changelog). Run the October 2 migration before it bites, and if you want Fable 5.1, enable its retention policy on purpose (Sept 1 changelog). Signups reopened September 3, so a team that was waitlisted can move now (Sept 3 changelog). Start from the Copilot plans page.

Enterprise on Enterprise ($39/user). Copilot was already the control-plane buy for GitHub-native orgs, and September widens the lead in governance rather than raw editing. You can now pin any supported model as the org or per-team default, content exclusions cover the app and CLI instead of just the IDE, and PR approval is scopeable by file path (Sept 2 default-model changelog, Sept 2 exclusions changelog, Sept 1 approval changelog). The action list is concrete: set the org default now that GPT-6 and Fable 5.1 are in the picker, enable Fable 5.1 retention deliberately, decide the PR-approval file paths, and complete the October 2 model swap. Re-check the Cursor comparison only if a team's day is genuinely editor-bound; for everyone whose work runs through GitHub, September keeps Copilot the standardize-here pick.