GitHub Copilot retired six models on September 1

Six models left GitHub Copilot on September 1. If your team has a favorite model pinned in Copilot settings, or an enterprise policy that allowlists specific ones, this is the week that quietly changed what your developers are actually writing code with.

What actually happened

GitHub's changelog lists the deprecations and their replacements: Claude Opus 4.5 and Opus 4.6 give way to Opus 4.7, 4.8 or 5. Claude Sonnet 4.5 and Sonnet 4.6 give way to Sonnet 5. Gemini 3.1 Pro is replaced by Gemini 3.7 Flash. Raptor Mini is replaced by MAI-Code-1.1-Flash.

The scope is wide. It covers Copilot Chat, inline edits, ask and agent modes, and code completions — effectively every surface. There is one exception: Claude Sonnet 4.6 remains available to individual Copilot subscribers on annual plans.

The administrative detail is the one that bites. GitHub notes the deprecated models are not automatically removed from systems, and that enterprise administrators should verify model availability in their Copilot policy settings and enable access to the replacements. If your organization allowlisted Opus 4.6 and nothing else, somebody has to go turn on its successor.

Why it matters for your business

A model swap inside a coding tool is not cosmetic. Prompt phrasing that worked reliably drifts. Agent mode behaves differently on long tasks. The Gemini 3.1 Pro to Gemini 3.7 Flash move in particular is a change in model class, not just a version bump — Flash-tier models are built for speed and cost, and if a team had Pro selected for a reason, that reason did not disappear.

Then check your bill. We have written before about how premium models bill inside Copilot, and a forced migration to a newer, more capable Opus is a migration to a different cost per request. Pull your usage report for the first week of September and compare it to late August before the month closes.

The durable lesson is the one every vendor keeps teaching: your model layer is rented, and it changes on someone else's schedule. Treat the specific model name as configuration, not architecture. Keep prompts and evaluation cases in your repo where you can rerun them against a new model in an afternoon. Keep a short regression suite — five real tasks from your codebase with known-good outcomes — so that "the new model is worse at our code" becomes a claim you can check instead of a hallway complaint. And put vendor deprecation notices somewhere a human reads them; this one shipped as a changelog entry on August 31 for a September 1 cutover.

Sources: GitHub Changelog: Selected GitHub Copilot models deprecated.