GitHub sounds the alarm: six Copilot models removed on December 14, Kimi K3 with no successor
🔎 One month to react, and one model left hanging
On November 13, 2026, GitHub published a changelog that deserves more than a skim: six models will be removed from Copilot on December 14, 2026, across all modes — Copilot Chat, inline edits, ask, agent, and code completions. Five of them have a named replacement. The sixth, Kimi K3, is listed as "to be announced": no successor, no date, no official plan B.
This is not routine maintenance. It is the third rotation of the Copilot catalog in three months, following the waves of October 2 and October 19, 2026. And it lands right in the middle of the end-of-year freeze period, when most teams freeze their deployments. If an agent workflow depends on a model being retired, the break will be silent: the model disappears from the selector, the replacement takes over, and nobody compares the results.
This article is an execution guide: the affected models and their successors, the critical case of Kimi K3, the policy checks admins need to perform, and a four-step audit method to run before the deadline. The facts come from the GitHub changelog, the recap from Today in AI (November 15, 2026), and analyses from WindowsForum and Josh Cook, Microsoft MVP.
The essentials
- Six models removed on December 14, 2026, across all Copilot modes: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3.
- Five named replacements: Claude Fable 5.2, Claude Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon, and Grok 4.7. Kimi K3: no replacement announced as of November 13, 2026.
- One notable exception: Claude Haiku 4.5 remains available for Copilot individual subscribers on annual plans.
- Enterprise/Business: suggested alternatives are enabled by default, unless an admin has disabled the global default or a specific model. Warning: some models remain disabled by default regardless of the policy.
- Third rotation in three months (October 2, October 19, December 14): model auditing should become a monthly ritual, not an emergency response.
- To do before December 14: pin your models, check the model policies, audit your agent workflows using the baseline → replacement → comparison method.
Recommended tools
| Tool | Main use | Price (November 2026) | Ideal for |
|---|---|---|---|
| GitHub Copilot | Agent, chat, inline edits and completions in the IDE and on GitHub.com | From $10/month (Pro); $19/user/month (Business) — check github.com, pricing evolves | Teams already committed, migration guided by model policies |
| Cursor | Multi-model agentic IDE | Around $20/month (Pro) — check cursor.com | Decoupling your agents from an imposed catalog |
| Cline | Open-source agent in VS Code, model of your choice (BYOK) | Free + API cost | Full control over the model and token budget |
| OpenRouter | Unified access to models for benchmarking before migration | Pay-as-you-go, free models available | Testing replacements without commitment |
| Hostinger | VPS for self-hosting open-weight models | From a few euros per month — check hostinger.com | Getting Kimi K3 out of GitHub's orbit by self-hosting Kimi K2.6 |
My take: if you're all-in on Copilot, stick with Copilot — but add an escape hatch. An OpenRouter account and a VPS cost less than a week of broken agent workflow in January.
The six models retired on December 14, and their replacements
On December 14, 2026, six models disappear from all Copilot modes; five are replaced by a named version, the sixth by nothing at all — for now.
Here is the exact mapping published in the GitHub changelog of November 13, 2026:
| Retired model | Official replacement | Point of attention |
|---|---|---|
| Claude Fable 5 | Claude Fable 5.2 | Direct migration |
| Claude Haiku 4.5 | Claude Haiku 5.5 | Haiku 4.5 remains accessible to individual subscribers on annual plans |
| Gemini 3.8 Flash | Gemini 4 Hydrogen | Direct migration |
| Gemini 4 Argon | Gemini 4 Carbon | Direct migration |
| Grok 4.6 | Grok 4.7 | Already the replacement for Grok 4.5 since October 19 |
| Kimi K3 | To be announced | No named successor as of November 13, 2026 |
The scope is broad: the deprecation applies to Copilot Chat, inline edits, ask mode, agent mode, and code completions. In other words, everywhere a model can be injected into your daily development workflow.
On the mechanics side, GitHub specifies that no action is required to remove the models after the date: the removal is handled platform-side. For Enterprise and Business, the suggested alternatives are enabled by default, unless an admin has disabled the global default or a specific model explicitly. Enabling and pinning is done via model policies in the Copilot settings.
The list is corroborated by the relays of the announcement, notably the Learnetto summary page and the Medium roundup, with the same blind spot on Kimi K3, listed without a successor.
The "enabled by default" trap
This is where many admins are going to get caught out. As WindowsForum points out, some models remain disabled by default regardless of policy — typically pre-GA and open-weight models, with Kimi K3 named explicitly.
In other words: "the replacement is enabled by default" in the changelog does not mean "the replacement is available in your organization." Between an enterprise policy that blocks open-weight models and a platform default that disables pre-GA versions, the path between the retired model and its successor can be severed without anyone noticing.
Check in your tenant, not in the changelog. That's rule number one of this migration.
Kimi K3 with no successor: the case that breaks workflows
If your agents run on Kimi K3, you have no official migration target: GitHub lists the replacement as "to be announced," with no name and no date.
The context makes this case particularly bitter. On October 2, 2026, GitHub deprecated Kimi K2.7 Code, designating Kimi K3 as the migration path. Teams that followed the official recommendation therefore find themselves, two months later, on a retired model with no safety net. This is the whiplash scenario: follow the vendor's advice, then pay the consequences.
The soundest advice comes from WindowsForum: don't plan around a replacement that GitHub hasn't announced. "To be announced" is not a roadmap, it's a non-answer. Building your migration on a ghost model means accepting that you'll migrate twice — or migrate in a panic.
Three concrete options are available to you:
- Migrate to a named model in the catalog and validate the choice with a comparative test (method detailed below). For agentic work, models like Claude Sonnet 4.6 (81.4 on agentic) or GPT-5.3 Codex (80) are among the current benchmarks, according to the June 2025 rankings.
- Leave the catalog and self-host. Moonshot AI's Kimi K2.6 remains available as open-weight (88.1 on agentic, June 2025 ranking): a VPS from Hostinger is enough to take back control of a model from the same family, without depending on a forced rotation.
- Switch families. Z.AI's GLM-5 (Reasoning), also self-hostable, is the closest open-weight alternative in terms of agentic profile (82, June 2025 ranking).
My opinion is firm: if Kimi K3 is critical to your setup, don't sit around waiting. An agent workflow that references a doomed model name with no successor isn't a plan, it's a time bomb with an expiration date: December 14.
Third Rotation in Three Months: Cadence is No Longer an Accident
December 14th will mark the third replacement of Copilot models in three months. This is no longer an incident; it is a cadence.
Let's put the waves in order, according to the changelog du 19 octobre 2026:
- October 2, 2026: Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code (replaced by Kimi K3), and Claude Opus 4.7 deprecated.
- October 19, 2026: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 deprecated, with GPT-6.1 Sol, GPT-6 Luna, and Grok 4.7 as replacements.
- December 14, 2026: the six models in this article.
Do the math: a Copilot agent workflow currently has a useful lifespan of about a quarter. Every model name hardcoded in a config, a system prompt, or a CI script is an expiring debt.
This instability is not unique to GitHub. Anthropic did the same thing with Claude Code, moving from free trials to credits on June 15th: the permanent free access is fading, models are rotating, and conditions are changing. The era of "I choose my model once and for all" is over.
The practical consequence: treat the model name as configuration, never as architecture. A model is replaced; an architecture survives replacements. If your workflow doesn't survive a change in a variable name, the problem is not GitHub's changelog.
Pin your models and check the policies: the admin express guide
Two checks before December 14: explicitly pin your models in the Copilot settings, and confirm that your planned replacements are actually enabled in your organization.
Here's the sequence I recommend, in five steps:
- Inventory references to the retired models: shared IDE configs, agent prompts, CI scripts, internal documentation. A search across the six names is enough to map out the risk landscape.
- Open the model policies in the Copilot settings and pin your choices. WindowsForum insists on this point: pinning your preferences prevents the platform default from deciding on your behalf.
- Check the default activations. On the Enterprise/Business side, the suggested alternatives are enabled by default unless explicitly disabled (global default or individual model). On individual accounts, test the model picker before the date, not after.
- Hunt down the models disabled by default regardless of the policy — pre-GA and open-weight versions, Kimi K3 leading the pack. If your planned replacement falls into this category, it won't show up on its own.
- Document the decision and schedule the next review. With a monthly rotation, a decision without a review date is as good as a forgotten one.
A point often overlooked for organizations: beyond the model choice, check the terms of use. Microsoft imposes undocumented fair-use caps on Copilot Chat for businesses — an "included" replacement that caps your agent consumption isn't a replacement, it's a bottleneck.
Audit your agent workflows in four steps
The basic method comes from Josh Cook, Microsoft MVP: save a baseline, run the replacement, compare the results. Expanded to four steps, it fits into an afternoon per workflow.
- Inventory: list every workflow that references one of the six retired models — agents, scripted inline edits, completions in CI, automated code reviews. This is the step everyone skips, and the one that costs the most when skipped.
- Baseline: capture the current model's outputs on a representative set of tasks — generation, refactoring, review, bug fixing. Without this capture, any subsequent comparison amounts to reading tea leaves.
- Replacement run: run the same task set on the official successor (Fable 5.2, Haiku 5.5, Gemini 4 Hydrogen or Carbon, Grok 4.7) — or on your substitute candidate if you're on Kimi K3.
- Comparison: output quality, latency, token consumption, tool-call failure rate. An agent that "works" but misses 10% more of your code conventions is a silent regression — the worst kind.
For teams whose agents rely on an internal knowledge base, one piece of architectural advice: decouple the context layer from the model. An approach like PageIndex, the vector-free RAG that reasons instead of searching lets the context layer survive model rotations — you swap the engine without rebuilding the gearbox.
Last point: keep the Learnetto summary page at hand as a checklist. It lists the December 2026 retirements and the status of Kimi K3, still without a successor.
What the migration will do to your bill
Replacements are newer and often more demanding; combined with usage-based billing, December 14th could shift your costs in either direction. Measure before you suffer.
The accounting context is as important as the technical context: GitHub Copilot has switched to token billing, ending the simple subscription in favor of a usage-based invoice. Under this regime, every model change alters the unit price of your tokens—and code completions are the most sensitive area, as they consume enormous volumes at low value per token.
Two cases deserve special attention:
- Inline completions: switching from Claude Haiku 4.5 to Claude Haiku 5.5 on millions of suggestions per month is multiplying a small price delta by a very large number of tokens.
- Agent in a loop: an agent mode that iterates on Gemini 4 Carbon rather than Gemini 4 Argon can consume more tokens per task, even at equal quality.
Good news for individuals: if you are an individual subscriber on an annual plan, Claude Haiku 4.5 remains available. This is a real grace period—use it to audit your workflows without pressure, not to postpone the decision.
And if your budget is tight, there is a way to benchmark without spending anything: free models via OpenRouter and Groq allow you to test alternatives without sacrificing quality. Testing before buying has never been easier—it has even become the only rational method in a catalog that changes every month.
Should you still build your stack on an imposed catalog?
Yes, but with your eyes open: Copilot remains a market leader, and the problem isn't quality — it's dependence on a catalog that rotates faster than your sprints.
Market figures prove the doomsayers wrong: according to Gartner's 2026 Magic Quadrant, OpenAI Codex, Cursor, and GitHub Copilot remain the leaders among enterprise coding agents. Copilot isn't a losing bet; it's a bet that demands discipline.
This discipline has a name: multi-model, with tested fallbacks. Concretely:
- Pin your models in each tool, and version those choices in your repo like any other configuration.
- Keep a fallback validated by the baseline → replacement → comparison method, not picked on a hunch the night of an outage.
- Diversify your entry points: an agentic IDE like Cursor or an open-source agent like Cline reduces the risk that a single catalog rotation paralyzes the entire team. Our roundup of the best AI tools for code details the options.
The rule I follow: no critical workflow should depend on a single model name. If a changelog can break your production pipeline, that's not a vendor problem — it's a design problem.
❌ Common Mistakes
Mistake 1: Waiting for the Kimi K3 replacement
What's wrong: "to be announced" has neither a name nor a date. Planning your Kimi K3 migration around a phantom successor means accepting a rushed migration, with no comparison, probably in January.
The solution: decide before December 14 between a named model from the catalog and a self-host of Kimi K2.6. Both can be tested in a single day; waiting, on the other hand, has no end date.
Mistake 2: Believing that "enabled by default" means "available to you"
What's wrong: some models remain disabled by default regardless of policy — pre-GA and open-weight versions, with Kimi K3 explicitly named by WindowsForum. The changelog describes the platform, not your organization.
The solution: check the activations in your tenant, via the model policies in the Copilot settings, and explicitly pin your choices.
Mistake 3: Auditing only Copilot Chat
What's wrong: the deprecation also affects inline edits, ask mode, agent mode, and code completions. Automated workflows and CI are the classic blind spots — and the most expensive ones, because nobody watches them run.
The solution: search for the six model names across all your repos, configs, and scripts, then treat every occurrence as a potential breaking point.
Mistake 4: Confusing the Haiku 4.5 exception with a general reprieve
What's wrong: only individual Copilot subscribers on an annual plan keep Claude Haiku 4.5 after December 14. Business and Enterprise teams have no equivalent — their Haiku 4.5 disappears just like the others.
The solution: if you fall under the exception, use it to audit at your leisure. Otherwise, integrate Haiku 5.5 into your migration plan right now.
Mistake 5: Migrating without a baseline
What's wrong: without capturing outputs before the migration, there's no way to know whether a regression comes from the model, the prompt, or your code. You'll discover the drifts through production bugs.
The solution: apply Josh Cook's method — baseline, run the replacement, compare — on a representative set of tasks, before the deadline.
❓ Frequently Asked Questions
What happens on December 14 if I do nothing?
The six models disappear from the selectors in all Copilot modes. On the Enterprise and Business side, the suggested replacements take over automatically, unless admin policy dictates otherwise; no action is required to remove the models. But nothing guarantees identical behavior: hence the audit before the date, especially for agent workflows.
Will Kimi K3 have a replacement?
GitHub lists "to be announced," with no name or date as of November 13, 2026. Neither the official changelog nor outlets like WindowsForum or Learnetto have announced a successor. In practice: don't plan around a ghost model. Test a named replacement from the catalog or prepare a self-host of Kimi K2.6.
Who keeps access to Claude Haiku 4.5 after December 14?
Only individual Copilot subscribers on an annual plan, according to the November 13, 2026 changelog. For everyone else — monthly Pro, Business, Enterprise — Haiku 4.5 disappears in favor of Haiku 5.5. If you qualify for the exception, it's a useful reprieve for auditing your workflows without deadline pressure.
Are replacements enabled automatically for everyone?
No. On the Enterprise and Business side, yes by default, unless an admin has disabled the global default or a specific model explicitly. On the individual accounts side, check the model selector yourself. One last caution: some models remain disabled by default regardless of policy, notably pre-GA and open-weight versions.
How do I test a replacement without breaking my workflows?
Capture a baseline on representative tasks, run the replacement on the same set, then compare quality, latency, token cost, and tool call failure rate. To broaden the test bench without spending anything, the free models from OpenRouter and Groq let you benchmark several candidates before committing to the migration.
Should you leave Copilot after this announcement?
No, not as a reflex: Gartner's 2026 Magic Quadrant places Copilot among the leaders in enterprise coding agents, alongside OpenAI Codex and Cursor. But the monthly rotation demands discipline: pinned models, tested fallback, recurring audits. It's diversification that protects your stack, not the exodus.
✅ Conclusion
On December 14, 2026, six Copilot models will disappear from all modes, and Kimi K3 still has no successor: you have one month left to pin your models, check your policies, and audit your workflows—otherwise, rotation will decide for you. For official details, the GitHub changelog of November 13, 2026 is authoritative.