The Butler
Back to briefings
GitHub ยท Model changes

Copilot's December Model Retirements: Test Your Replacement Before the Deadline

· The Butler · 4 minute read

The metallic blue-eyed robot Butler standing ready in a warm library, wearing a tuxedo and top hat.
Butler editorial illustration, not a product screenshot.

If your team has found a Copilot model that works reliably, put December 14 on the calendar. GitHub has announced another set of model retirements. The useful response is not to chase every new model. It is to check which everyday jobs depend on the ones leaving.

The November 13 notice covers selected models across GitHub Copilot experiences. This is a Copilot availability change, not an announcement that those models are disappearing from every provider or application. GitHub's retirement notice

What is changing

GitHub lists these December 14 replacements: Claude Fable 5 moves to Fable 5.2; Claude Haiku 4.5 to Haiku 5.5; Gemini 3.8 Flash to Gemini 4 Hydrogen; Gemini 4 Argon to Gemini 4 Carbon; and Grok 4.6 to Grok 4.7. Kimi K3's replacement remains to be determined. Annual individual subscribers retain access to Haiku 4.5 under the stated exception. Check the announcement against your subscription rather than treating the list as identical for everyone. Affected models and exception

That gives you a month to test a replacement while the old choice is still available. A newer label does not establish that it will follow your repository's conventions equally well.

Make a small, representative test pack

Choose three tasks you actually repeat. For example: explain an unfamiliar function, repair a small bug with an existing test, and draft a change that must follow your team's style. Use a branch and non-sensitive examples appropriate for your organization's policy.

Save the starting state, the instructions and the checks that define a good result. Run the same task with your current choice and the proposed replacement. Compare the work, not just the fluency of the explanation.

Did the patch pass the tests? Did it modify files outside the requested scope? Did you spend longer correcting it? A five-minute answer that needs twenty minutes of repair is not necessarily an improvement over a slower, dependable answer. Our earlier guide to measuring one AI workflow offers a useful way to count that review work.

Check the client and the policy

GitHub's supported-model reference distinguishes availability by plan and client and lists version requirements for some experiences. Check the editor and extension your team actually uses. Do not assume a model visible to one colleague will appear everywhere. Supported models and requirements

For Business and Enterprise, GitHub says alternatives are enabled automatically unless the global default is off or the model has been disabled. In those cases, administrators need to review model policies. Retired models do not require manual removal. Administrator details

An automatic replacement is a convenience, not a reason to broaden permissions indiscriminately. Keep your organization's existing data and access rules in place. Record the tested model and client version alongside the result so a later change is easier to diagnose.

Leave yourself a clear handover

Before the deadline, write a short note for your team: what is changing, which replacement you tested, which tasks passed, and where people should report regressions. If you depend on Kimi K3, do not invent a promised successor; keep watching the official notice and test an available, permitted alternative against your own work.

The goal is an uneventful December morning. You want developers thinking about the change they are building, not discovering midway through it that their familiar model has gone.

Sources checked November 14, 2026. This is migration planning, not a benchmark or a claim that we tested these models.