1. A concrete deadline across Copilot experiences
The GitHub changelog announcement says the selected models will be deprecated on December 14, 2026, across Copilot experiences, including Copilot Chat, inline edits, ask and agent modes, and code completions. GitHub asks users to update workflows and integrations before that date.
The scope makes this more than a reminder to change a preferred chat model. Our editorial recommendation is to review each place your team deliberately selects an affected model, including repeatable instructions and integration configuration. Record the client, task, and model choice together so that a successful chat experiment does not become your only evidence for a different coding workflow.
GitHub says no action is required to remove the models after deprecation. That statement concerns removal; it does not establish that your chosen replacement will produce acceptable results on existing tasks.
2. Five suggested alternatives, one unresolved replacement
GitHub lists the following retirement schedule and suggested alternatives. These are vendor recommendations, not independently measured equivalence claims.
| Retiring model | Retirement date | GitHub’s suggested alternative |
|---|---|---|
| Claude Fable 5 | December 14, 2026 | Claude Fable 5.2 |
| Claude Haiku 4.5 | December 14, 2026* | Claude Haiku 5.5 |
| Gemini 3.8 Flash | December 14, 2026 | Gemini 4 Hydrogen |
| Gemini 4 Argon | December 14, 2026 | Gemini 4 Carbon |
| Grok 4.6 | December 14, 2026 | Grok 4.7 |
| Kimi K3 | December 14, 2026 | To be announced |
*GitHub explicitly exempts individual Copilot subscribers on annual plans from the Claude Haiku 4.5 retirement. The announcement says that model will remain available to those subscribers. It does not give an end date for the exception or extend it to the other five models.
For an affected Kimi K3 workflow, our recommendation is to evaluate an available alternative now while monitoring GitHub’s announcement for a named successor. Do not build a migration plan around an assumed Kimi release or replacement date.
3. Verify access in the client you actually use
GitHub’s supported-model documentation says availability depends on both your Copilot plan and where you use Copilot. It also lists minimum client versions for some models, while warning that the version information is tentative. A model appearing in the overall catalog therefore does not establish availability in every developer’s setup.
The retirement announcement says suggested alternatives are automatically enabled for Business and Enterprise customers under default model enablement, unless an administrator has disabled the global default or the particular model. However, the supported-model documentation separately excludes Claude Fable models, including Fable 5.2, from default enablement. These statements need a qualification: do not assume every suggested alternative will appear automatically, especially Fable 5.2.
As a proposed check, confirm the candidate appears in the intended client and can handle the intended interaction before scheduling the migration. Keep an unavailable candidate separate from one that runs but performs poorly; they require different follow-up.
4. Compare replacements on familiar work
Our proposed evaluation starts with three bounded tasks from your own repository: explaining an unfamiliar function, fixing a reproducible bug, and making a small change across related files. Choose tasks with outcomes you can inspect. The purpose is to reveal differences that matter to your work, not to establish a general model ranking.
For each candidate, use the same starting revision, instructions, and relevant context. Record the model name, client version, changes produced, checks completed, and human corrections required. If the retiring model remains available to your account, run it against the same tasks to create a comparison before the deadline.
Define acceptance before reading the output. A bug fix should address the reproduction and pass relevant regression checks; a refactor should preserve the behavior you intended to keep. Our guide to regression tests for AI-generated code explains how to make that evidence useful.
These are proposed checks, not results from testing the listed replacements. GitHub’s replacement table provides a starting point for selection, but it does not establish equivalent instruction following, code quality, or review effort.
5. Finish with a workflow decision
For a small team, the useful migration record can be brief: affected task, retiring model, accepted replacement, client used, supporting checks, and unresolved limitations. Assign someone to resolve each remaining dependency before December 14. A solo developer can use the same record to avoid rediscovering why a model was chosen.
Move a workflow only after the candidate meets its acceptance criteria. Keep human review and existing checks in place during the transition. If a replacement changes more files than requested or needs repeated correction, narrow the task and reassess it. Our coding-assistant evaluation guide offers additional criteria for judging that fit.
The annual Haiku exception may remove immediate migration pressure for eligible individuals, but it should not be assumed to cover collaborators with different subscriptions. Kimi users still have an unanswered replacement question. For everyone affected, the concrete deliverable is a supported model choice backed by checks on the workflow that must continue after the cutoff.
Sources & further reading
Primary references used for product and technical information. Features and terms can change; check the provider before committing.
- GitHub Changelog: Upcoming deprecation of selected GitHub Copilot models in DecemberAccessed 2026-11-14
- GitHub Docs: Supported AI models in GitHub CopilotAccessed 2026-11-14