GitHub Copilot retires six models on September 1, and every Copilot surface is in scope

GitHub is retiring six models from every Copilot experience on September 1, 2026, and any workflow or integration that pins one of them stops routing to it that day. The July 31 changelog lists the models, names a suggested replacement for each, and carves out one narrow exception. Enterprise administrators are on the hook to enable the successor models through Copilot settings before the cutoff.

The six models on the retirement list

Per the changelog, the deprecations span three vendors:

• Gemini 3.1 Pro is retired, with Gemini 3.6 Flash as the suggested alternative.
• Claude Opus 4.5 and Claude Opus 4.6 are both out; the suggested paths forward are Claude Opus 4.7, 4.8 or 5.
• Claude Sonnet 4.5 goes, with Claude Sonnet 5 as the target.
• Claude Sonnet 4.6 also goes for most tiers, again pointing at Claude Sonnet 5.
• Raptor Mini is retired, with MAI-Code-1-Flash as the alternative.

The changelog does not say whether the recommended successors match the deprecated models on latency, token budget or cost. They are the family GitHub is pointing users at, and no more than that.

Where the deprecation lands

The change covers "all GitHub Copilot experiences," and the changelog spells them out: Copilot Chat, inline edits, ask and agent modes, and code completions. For a CI/CD team, that reaches every place a Copilot-driven action can enter a pipeline. If a repository's automation flows through a Copilot agent, a chat surface, an inline-edit call or a code-completion path that names one of the six models, the reference stops resolving on September 1.

The exception is narrow. Claude Sonnet 4.6 remains available to individual GitHub Copilot subscribers on annual plans, so that tier retains a Sonnet offering. Org and enterprise plans do not get the carve-out.

What to change, and by when

GitHub's instruction to users is short: update your workflows and integrations to use a supported model before the deadline. For an enterprise, that reads as an audit rather than a one-line edit. Any workflow, custom action, Copilot Extension configuration or agent policy that names a specific model has to be found, updated, and reviewed. The changelog adds a second task for enterprise administrators: alternative models must be enabled through Copilot settings for the successors to be reachable inside the tenant. After the deadline, GitHub says no cleanup is required to remove deprecated models from configuration; the platform stops offering them.

The residual caveat is the one the changelog does not spell out. A workflow that names Claude Opus 4.5 today and gets pointed at Claude Opus 5 next month is not the same workflow. Prompt behaviour, refusal patterns and token cost can all move on a model swap, and Copilot integrations that read like tested software should be treated that way: migrate on a branch, run the suite, and cut the change over before September 1, not on it.

How model retirement usually shakes out at this cadence

A per-model sunset date, a suggested replacement, and no promise of behaviour parity is the pattern most platforms hosting third-party LLMs have converged on. What is worth noting here is the density: six models across three families in one window, with no staggering on the published list. A shop running multiple Copilot surfaces has the migration work concentrated in the same month rather than spread out.

The clock is short. GitHub filed the notice on July 31; the cutoff is September 1. Anything that has to ship before then ships against a moving model backend.