GitHub Copilot Is Retiring These AI Models on October 19, 2026, Here's What to Switch To

On September 18, 2026, GitHub announced that six Copilot models will be retired on October 19, 2026, twelve days from now. The shutdown covers every Copilot surface: Chat, inline edits, ask and agent modes, and code completions. If your team has pinned a model anywhere, a policy, a config file, a runbook, an evaluation harness, that pin stops working on the 19th.

This is the plain-English version of GitHub's changelog notice: which models are leaving, the official replacements, what actually breaks, and the checklist to get your team migrated before the deadline. If you are thinking about the security side of AI coding tools more broadly, see our companion piece at AI coding agent security risks.

Key takeaways

• Six models retire October 19, 2026: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5.
• GitHub's suggested replacements: Gemini 3.8 Flash, GPT-5.6 Sol (for the full-size GPT-5.x models), GPT-5.6 Luna (for the minis), and Grok 4.6.
• An earlier wave already retired October 2: Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7.
• Under default model enablement, replacements are automatically enabled for Copilot Business and Enterprise, unless an admin disabled the global default or the specific model.
• Watch the cost impact: GPT-5.4's replacement moves up a class on GitHub's rate card (Versatile to Powerful), which raises per-token pricing.

1. The six models retiring October 19

From GitHub's September 18 changelog, here is the complete deprecation table with GitHub's suggested alternative for each:

GitHub's guidance is direct: update workflows and integrations to the supported models before the date. No action is required to remove the retired models afterward, they simply stop being available.

The naming deserves a decoder ring. In the GPT-5.x family, "Sol" is the full-size model aimed at interactive and agentic coding work, while "Luna" is the smaller, faster, cheaper sibling for lightweight jobs. So the migration is: full-size GPT-5.x models go to Sol, minis go to Luna. The Gemini and Grok swaps are straightforward version bumps.

2. The wave you may have already missed: October 2

October 19 is actually the second retirement wave this month. On October 2, 2026, GitHub retired four more models:

If your team already migrated off these, you have the pattern down. If you are only hearing about the October 2 wave now, check whether anything still references those names, they are already gone, and any workflow pointing at them is already broken.

3. What actually breaks

The retirement applies across all Copilot experiences, so the blast radius depends on how deeply your team customized model selection:

• Chat model picker: anyone who manually selected a retiring model will find it missing after October 19. The picker will show the enabled replacements instead.
• Agent mode and ask mode: agentic workflows pinned to a specific model (in settings, workspace configs, or custom agents) will fail or fall back when the model disappears.
• Code completions: the default completion model is managed by GitHub, but teams that pinned completions to a named model need to re-pin.
• The long tail: this is where teams get bitten. Model names hide in CI configs, evaluation harnesses, internal runbooks, wrapper scripts, prompt libraries, and team onboarding docs. Every reference to the six retiring names is a small landmine.

One thing that does not break: GitHub says no action is required to remove the models after deprecation. The risk is entirely on the side of things that explicitly name the old models.

4. The default-enablement nuance admins must check

GitHub's most misunderstood line in the announcement: under default model enablement, the suggested replacements are automatically enabled for Copilot Business and Copilot Enterprise customers, unless an administrator has turned off the global default or explicitly disabled the replacement model.

That "unless" carries the whole operational risk. If your org disabled the global default (common in regulated industries and companies with approved-model lists), the replacements will not appear for your users automatically. Someone with admin rights must enable them through model policies in Copilot settings. And if an admin explicitly disabled, say, all Gemini models for policy reasons, Gemini 3.8 Flash stays off and the team needs a different migration path.

Action for admins: open Copilot settings now, check the model policy list, and confirm each replacement model is enabled for the teams that need it. Do not assume the automatic path applied to you.

5. The team migration checklist

Treat this like a dependency migration, not a settings tweak. Twelve days is enough if you start now.

1. Inventory pinned models. Search your repos, configs, and docs for the six retiring names: GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, Grok 4.5, plus the four October 2 names if you never checked. Include evaluation baselines and prompt regression suites.
2. Check admin policies. Confirm the replacement models are enabled in Copilot settings for Business/Enterprise, especially if your org turned off the global default.
3. Test replacements on real work. Suggested does not mean identical. Run your representative coding, review, and agent tasks against the replacement before the cutoff and compare behaviour, particularly for agent-mode workflows, where model differences compound over long runs.
4. Watch the price changes. Two swaps move up a class on GitHub's rate card: GPT-5.4 goes from Versatile to Powerful (its replacement costs more per token), and Kimi K2.7 Code to Kimi K3 also steps up. If you budget Copilot spend per team, update the forecast. The Gemini and Grok swaps hold their class.
5. Check minimum plans. Access varies by tier: GPT-5.6 Sol requires Pro+ or above (Pro, Max, Business, Enterprise get Luna); the Gemini and Grok replacements are available on Pro and above. If anyone on your team is on base Pro, confirm their replacement is actually in their picker.
6. Update the docs. Runbooks, onboarding guides, and internal wikis that name the old models should point at the replacements, with the migration date noted.
7. Put model selection behind config. The deeper lesson of 2026's repeated retirement waves: hard-coding model names into workflows guarantees future fire drills. A configuration layer with a recorded replacement, date, and benchmark set turns the next wave into a routine change.

6. Why this keeps happening

This is GitHub's third Copilot model retirement wave in under two months, and it will not be the last. Copilot has become a multi-model platform where the model underneath a workflow can change independently of the tool around it, which makes model selection, policy, and migration look increasingly like dependency management.

The practical upshot: schedule a recurring review of your pinned models (quarterly is plenty), keep that configuration layer from step 7, and expect GitHub Universe (October 28–29, just after the cutoff) to bring a fresh round of model announcements. The teams that treat models as pinned dependencies with a migration process will barely notice; the teams that treat them as permanent features will keep getting surprised.

Practical next steps

• Today: search your codebase and docs for all ten retired model names (both October waves).
• This week: verify replacement models are enabled in Copilot admin settings; run your key workflows against the replacements.
• Before October 19: re-pin Chat, agent mode, and completions to supported models; update runbooks.
• After October 19: confirm nothing references the retired names; note the migration in your team changelog.
• October 28–29: watch GitHub Universe announcements for the next model roster changes, and fold them into your quarterly review.

The bottom line

Six models retire October 19 across every Copilot surface, and the migration is straightforward if you do it deliberately: Sol for full-size GPT-5.x, Luna for the minis, 3.8 Flash for Gemini, 4.6 for Grok. The teams at risk are not the ones using defaults, GitHub handles those, but the ones with pinned models, disabled defaults, or model names scattered through configs. Twelve days, one checklist, no surprises.

Sources

• GitHub Changelog, Upcoming deprecation of selected GitHub Copilot models in mid-October (Sep 18, 2026)
• C# Corner, Preparing for GitHub Copilot Model Changes
• LinkLoot, GitHub Copilot sets October 19 deadline for six model migrations
• ByteIota, GitHub Copilot Drops GPT-5.5 and Grok 4.5 on October 19
• HotTea, GitHub will retire six Copilot models on October 19
• Pondero, GitHub Copilot deprecated models, October 2026: what to switch to and what it costs