Copilot Model Migration: A 30-Day Checklist

A Copilot model migration is due on December 14, 2026, when GitHub removes six models from every Copilot experience, including chat, inline edits, ask and agent modes, and code completions. Find pinned model IDs in editors, the Copilot SDK, workflows, and organization policies before a call still names a model GitHub will not serve. EpiFive counted 30 days from its November 14 reading of the November 13 changelog to that cutoff.

Updated November 15, 2026

Which models leave on December 14?

GitHub’s December deprecation notice names Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3. Claude Haiku 4.5 remains only for individual annual-plan subscribers. Kimi K3 has no named successor; the announcement says the replacement is “to be announced,” so choose another supported model now. Copilot itself keeps working; only the selectable list shrinks.

How do you run a Copilot model migration?

GitHub asks teams to update workflows and integrations before the date. It publishes no usage report, and it does not say what a call to a missing model will do. Work through the surfaces that can still name one.

1. List which of the six models people use. Ask in chat and read shared docs; the announcement describes no usage report.
2. Search custom agents, workflows, and Copilot SDK integrations for a retired name. SFEIR flags those configs because fallback behavior is undocumented.
3. Click your profile picture, choose Organizations, then Settings. Under “Code, planning, and automation,” click Copilot, then Models. If an enterprise owner has already set a model policy, you cannot override that setting at the organization level.
4. Under default model enablement, suggested alternatives turn on for Business and Enterprise unless an admin disabled the global default or that model. If the default is off, enable them in model policies or the picker offers fewer choices after December 14.
5. Confirm Gemini 4 Hydrogen for Flash-pinned agents and the SDK. Rollout is gradual. If global default enablement is off and teams used Argon, SFEIR advises enabling Gemini 4 Carbon too.
6. In VS Code, click the chat icon in the title bar, then the CURRENT-MODEL dropdown at the bottom of the chat view. Visual Studio, JetBrains, and Xcode open chat their own way, then use a CURRENT-MODEL dropdown. In Eclipse, click the current model at the bottom right of the chat panel. Multi-model chat in Xcode requires the GitHub Copilot for Xcode extension. That page did not list GitHub.com steps.
7. Rewrite conventions that name a deprecated model. On or after December 14, confirm the picker shows only supported models. No one has to delete the retired models after the date.

What should admins verify before the cutoff?

CASRAI treats a suggested replacement, an enabled policy, and approval for a use as separate decisions. Default policy covers eligible generally available models admins have not configured, except models outside GitHub’s data-retention agreement.

Claude Fable 5.2 is off by default, per its November 9 notice. Anthropic retains prompts and outputs for safety classifiers by default, and a zero-data-retention exemption through the end of 2026 still needs its own enablement step.

Business organizations must allow members to switch models. Policies beyond the basic models may cost more, and the changelog does not say a replacement draws the same allowance. If compliance restricts providers, review each replacement first; the notice says nothing about training. Research teams should record the selected model with the revision so a generator change is not mistaken for a software change.

What does the Flash-to-Hydrogen price jump change?

Since June 2026, chat, agent mode, code review, and Copilot CLI spend AI Credits, with 1 credit equal to 0,01 $. Code completions stay off that meter even though completions lose these models too.

SFEIR, citing GitHub’s pricing page on November 15, says an uncached Hydrogen token costs twice 3.8 Flash’s promotional rate, while a cached read stays at 0,075 $ per million tokens. Hydrogen is on Pro, Pro+, Max, Business, and Enterprise, not the free plan. As of that date it was absent from automatic selection, where 3.8 Flash was the only Gemini model, and GitHub has not said it will replace Flash there.

SFEIR recalculated a November 14 Data Studios scenario: 20 calls, 30 000 input tokens, 70 percent cache reads, and 1 500 output tokens each. The sample costs 0,5265 $ (52,65 credits) on Hydrogen versus 0,279 $ (27,9 credits) on 3.8 Flash. It is a simulation: cache rate changes the total, and GitHub does not list Hydrogen among Copilot models whose reasoning level you can set.

For engineering managers, that simulation means the same Business envelope covers almost half as many agent sessions, so an added usage budget is what keeps people unblocked at month end. Compare the models on about twenty real tasks per team and log success rate, iterations, and credits. Vendor scores were not measured in Copilot. Switch steps are in our Gemini 4 Hydrogen migration checklist; seat math is in coding-agent budgets for Carbon and Fable 5.2.