GitHub Copilot Retires Six Models Today: The Great September 1 Model Purge

Six models disappeared from GitHub Copilot today. If you opened the model picker in VS Code or on github.com this morning and found your favorite missing, it is not a bug: September 1, 2026 is the deprecation date GitHub announced a month ago in its changelog, and it applies “across all GitHub Copilot experiences” — Copilot Chat, inline edits, ask mode, agent mode, and code completions.

The retired list spans three vendors: Google’s Gemini 3.1 Pro, Anthropic’s Claude Opus 4.5, Claude Opus 4.6, Claude Sonnet 4.5, and Claude Sonnet 4.6, plus Microsoft’s own in-house completion model, Raptor Mini. The Claude 4.x generation is now effectively gone from the world’s most widely deployed AI coding assistant, and the replacements GitHub suggests tell an interesting story about where Copilot’s model strategy is heading.

The full retirement map

Per GitHub’s official July 31 changelog, the mappings are:

Retired model	Suggested alternative
Gemini 3.1 Pro	Gemini 3.6 Flash
Claude Opus 4.5	Claude Opus 4.7, 4.8, or Opus 5
Claude Opus 4.6	Claude Opus 4.7, 4.8, or Opus 5
Claude Sonnet 4.5	Claude Sonnet 5
Claude Sonnet 4.6	Claude Sonnet 5
Raptor Mini	MAI-Code-1-Flash

Each mapping deserves a closer look, because none of them are simple like-for-like swaps.

Gemini 3.1 Pro → Gemini 3.6 Flash. This is the most asymmetric replacement on the list: a Pro-tier model being retired in favor of a Flash-tier one. Google’s positioning is that the current Flash generation has caught up to (or surpassed) the previous Pro generation on coding-adjacent benchmarks, at a fraction of the latency and cost. Developers who relied on Gemini 3.1 Pro for heavy reasoning in agent mode may want to A/B test the swap before committing — or look at Gemini’s current Pro tier, which remains available in Copilot separately.

The Claude 4.x exodus. Four of the six retirements are Anthropic models, and together they end the Claude 4 era in Copilot. The suggested ladder is telling: Opus 4.5 and 4.6 users are pointed at Opus 4.7, 4.8, or Opus 5 — and since Claude Opus 5 landed in Copilot on July 24 for Pro+, Max, Business, and Enterprise tiers, the writing on the wall is that GitHub wants Opus traffic consolidating on the frontier model. Sonnet users move to Claude Sonnet 5, which Anthropic shipped June 30 with performance the company says approaches Opus 4.8 at $3/$15 per million tokens standard pricing.

Raptor Mini → MAI-Code-1-Flash. The least headline-friendly change may be the most strategically significant. Raptor Mini was Microsoft’s in-house model for Copilot’s free/low-tier experiences. Its replacement, MAI-Code-1-Flash, is Microsoft AI’s first purpose-built coding model: a 137B-parameter mixture-of-experts with a 256K context window, priced at $0.75/$4.50 per million tokens, and generally available across Copilot Business and Enterprise since late June. Retiring Raptor Mini in its favor completes the transition of Copilot’s completions engine to Microsoft’s own model family — a meaningful step away from dependence on OpenAI for the baseline experience.

The annual-plan exception

Buried in the changelog is a carve-out worth knowing: Claude Sonnet 4.6 will remain available to individual Copilot subscribers on annual plans, “so they retain a Sonnet offering.” The deprecation explicitly does not apply to those customers.

It is a small clause with real implications. Individual annual subscribers — many of whom chose that plan specifically to lock in model access and pricing — keep a model that everyone else loses today. Everyone else, including monthly individual subscribers and all business tiers, must move to Sonnet 5. If your team has a mix of plan types, your developers may now be looking at different model menus, which is worth clarifying internally before someone files a “my model vanished” bug.

What admins actually need to do today

GitHub’s guidance is straightforward but easy to miss:

1. Update workflows and integrations that pin specific model names. Anything hardcoded to claude-4.5-sonnet-style identifiers or Gemini 3.1 Pro will start failing silently or loudly, depending on how it handles model-not-found errors.
2. Copilot Enterprise administrators may need to enable the alternative models through model policies in Copilot settings. The replacements are not automatically visible: an admin must enable the policy for each specific model, then verify it appears in the Copilot Chat model selector in VS Code and on github.com.
3. No action is required to remove the old models — deprecation handles that side automatically.

The admin-policy step is the classic gotcha. A developer asks “where did Opus 4.6 go?”, the answer is “enable Opus 5 in the org policy,” and if nobody knows that, the ticket bounces between engineering and IT for a week.

Why this purge matters beyond Copilot

Set against the rest of this week’s news, the September 1 purge fits a pattern: the AI industry has entered a phase of aggressive generation-sunsetting, and the Claude 4.x removal is its largest single stroke yet.

Yesterday alone, three other deadlines landed — Claude Sonnet 5’s introductory $2/$10 pricing expired, GPT-5.4 left Codex for ChatGPT-authenticated users, and Moonshot fully sunset kimi-k2.5 and moonshot-v1. Model lifecycles that used to run 18-24 months are compressing to 6-9. When four Claude models vanish from the single most-deployed coding assistant on one day, “build for model portability” stops being architecture-astronaut advice and becomes basic operational hygiene.

There is also a consolidation signal here. Copilot’s suggested replacements lean conspicuously toward the newest generation of each family — Opus 5, Sonnet 5, Gemini 3.6 Flash, MAI-Code-1-Flash — and Microsoft’s own model now covers the completions baseline. GitHub is not just clearing old inventory; it is steering the world’s developers onto a narrower, newer, partly first-party model set. That is cheaper to serve, easier to evaluate, and strategically safer for Microsoft.

For developers, the practical takeaway is simple: check your defaults, check your pinned model strings, and if you are an Enterprise shop, check your model policies today. The six models are gone; the map above is where they went.