GitHub Copilot retires six models: two need a manual swap
Six Copilot models retire December 14. Most replacements switch on through the default policy, but Fable 5.2 needs an admin and Kimi K3 has no replacement yet.
Sean McLellan
Lead Architect & Founder

GitHub will deprecate six models across every Copilot experience on December 14, 2026: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. The December deprecation notice, posted November 13, names a suggested alternative to five of them. Business and Enterprise admins now have to work out which swaps happen on their own and which need someone to act before that date.
Four of the five alternatives switch on automatically, unless an admin has turned off automatic enablement or blocked that model. That leaves two of the six retirements that need a person. Fable 5.2 sits outside the automatic path and has to be switched on by hand, with a data-retention acknowledgment. Kimi K3 has no named successor yet. The swaps also change what each request costs, in both directions.
Which GitHub Copilot models are being deprecated?
Six models, all on December 14, 2026, across chat, inline edits, ask and agent modes, and code completions. The notice pairs them like this:
Scroll sideways to see all 3 columns.
| Retiring model | Suggested alternative | Enabled automatically? |
|---|---|---|
| Claude Fable 5 | Fable 5.2 | No, an admin must enable it |
| Claude Haiku 4.5 | Haiku 5.5 | Yes |
| Gemini 3.8 Flash | 4 Hydrogen | Yes |
| Gemini 4 Argon | 4 Carbon | Yes |
| Grok 4.6 | Grok 4.7 | Yes |
| Kimi K3 | To be announced | No alternative yet |
There is one exception: Haiku 4.5 stays available to individual subscribers on annual plans. Admins don't need to do anything to remove the retired models. The notice asks customers to move workflows and integrations to supported models before that date.
This is the third batch since the start of October. Our post on the October 2 Copilot model retirement covered the first, when Kimi K3 was the suggested successor to Kimi K2.7 Code. Now K3 is leaving too, about ten weeks later.

Four swaps happen automatically unless an admin opted out
Haiku 5.5, Hydrogen, Carbon and Grok 4.7 are switched on in Business and Enterprise organizations that leave automatic enablement alone. The notice says the suggested alternatives are enabled "unless an administrator has turned off the global default or explicitly disabled the model."
The default availability docs explain the mechanism. A model nobody has configured carries the label "Delegate to Default Policy" and inherits the enterprise or organization setting. A model set to Enabled or Disabled keeps that choice. So if your enterprise has the setting at Disabled, or someone blocked a successor when it launched, nothing takes the old model's place on December 14 until an admin steps in. Our earlier post on one Google launch that arrived switched on has a worked example.
Plan coverage differs too. Hydrogen reaches individual Pro subscribers. The 5.2 note and Google's November 10 note list Pro+, Max, Business and Enterprise, without Pro.
Fable 5.2 has to be switched on by hand
This is the one suggested alternative that automatic enablement can't reach. The notice's wording reads as if it covers every alternative, but the availability docs exclude all three versions (5, 5.1 and 5.2) because they are "not covered by GitHub's data retention agreement," and say they stay disabled "regardless of your 'Default availability' policy setting."
Teams on Fable 5 today already made this choice once, since that version needed the same manual step. The newer version has its own setting, though, and its launch note says turning it on "constitutes acknowledgment" that Anthropic retains prompts and outputs to run safety classifiers, data that isn't used in training. Eligible customers can ask their account team about zero data retention through the end of 2026. Our post on the 5.2 retention opt-in walks through that choice.
If nobody acts, those users are left with whatever else is enabled on December 14. That can include Fable 5.1, which is not on the retirement list.
Kimi K3 has no replacement yet
The notice lists Kimi K3's alternative as "To be announced." The availability docs also place open-weight models, naming DeepSeek and Kimi K3, outside automatic enablement, so anyone using K3 today had it turned on deliberately. BaristaLabs interpretation: move that work to another enabled model now rather than wait on an announcement that may not arrive before the deadline.
How do the replacement prices compare?
Two swaps cost less per token, two stay flat at the listed rates, and Hydrogen costs more. The Argon replacement's listed rates rise on December 1. These models bill at provider list price in AI Credits, at $0.01 per credit, once usage passes the plan's included allowance. Rates below come from the models and pricing page, per 1 million tokens, standard tier:
Scroll sideways to see all 4 columns.
| Swap | Input | Cached input | Output |
|---|---|---|---|
| Fable 5 to 5.2 | $10.00 to $8.00 | $1.00 to $0.20 | $50.00 to $40.00 |
| Haiku 4.5 to 5.5 (up to 100K input) | $1.00 to $0.10 | $0.10 to $0.01 | $5.00 to $0.50 |
| 3.8 Flash to Hydrogen | $0.75 to $1.50 | $0.075 to $0.075 | $3.75 to $7.50 |
| Argon to Carbon | $2.00 to $2.00; successor $4.00 from Dec. 1 | $0.10 to $0.10; successor $0.20 | $10.00 to $10.00; successor $20.00 |
| Grok 4.6 to 4.7 (up to 200K input) | $2.00 to $2.00 | $0.50 to $0.50 | $6.00 to $6.00 |
A few details sit behind those numbers:
- 3.8 Flash carries promotional pricing through December 31, 2026, but it leaves the service on December 14. The same Flash-class traffic on Hydrogen costs twice as much on uncached input and output. Our Hydrogen vs 3.8 Flash comparison covers the model differences.
- The Argon swap is $2 input and $10 output on both models at the listed promotional rates. The successor's promotion ends December 1, when its rates rise to $4 input and $20 output. Argon's promotion also ends that day, and the pricing page lists no later Argon price.
- Haiku 5.5 costs a tenth of Haiku 4.5 below 100,000 input tokens, and $0.50 input and $2.50 output above that. The pricing page files 5.5 under "Lightweight" and 4.5 under "Versatile," so check output quality on the tasks the older one handled.
- Code completions aren't billed in AI credits, so these rates apply to chat and agent work, not completions.
BaristaLabs arithmetic: a team that spends 500 AI credits ($5) a month on uncached 3.8 Flash requests would spend about 1,000 credits ($10) on Hydrogen at the same token counts. The $5 difference is small, but it recurs each month and adds up across teams that share a pooled organization allowance.
Retirements now arrive every few weeks
The December batch follows two in October. The supported-models page's retirement history includes Claude Opus 4.7 on October 2, then Gemini 3.7 Flash, GPT-5 mini, GPT-5.4, GPT-5.4 mini and GPT-5.5 on October 19. Each of those entries points to the latest release in its line, such as the latest GPT Luna or Sol, rather than a fixed version.
BaristaLabs interpretation: a team guide that says "use Haiku 4.5 on quick refactors" goes stale on a schedule the team doesn't control. Writing the guide by tier (a fast, cheap option on small edits; a stronger one on long agent runs) and keeping a short mapping from tier to current name means each retirement becomes a one-line edit instead of a rewrite.
Who needs to act before December 14
BaristaLabs interpretation: most organizations that leave automatic enablement on will see four swaps happen without any work. Three groups should check their settings now:
- Organizations that turned automatic enablement off, or blocked any of the four automatic successors. Without a change, people on the retiring models end up with no direct replacement in the picker.
- Teams on Fable 5. Someone has to make the 5.2 retention decision and switch it on, or pick another option.
- Anyone with Kimi K3 or a retiring name in scripts, agent configuration or SDK calls. The picker change doesn't touch those files. Search repositories and shared settings to find the six names, and swap each hit deliberately.
Recommended first step: list your model settings
BaristaLabs recommendation: before December 1, open your enterprise or organization model settings and write down the status of the six retiring models and their five alternatives: Enabled, Disabled or Delegate to Default Policy. An alternative marked Disabled, or delegated while your enterprise setting is Disabled, is a gap if its predecessor is in use. Close those gaps, record the 5.2 retention decision in writing, and pick a stand-in to cover Kimi K3. If you manage these settings through the API, our post on change control with the model-policy API shows how to keep the change reviewable. Doing this by December 1 also means you see Carbon's higher price on real usage before the deadline rather than after it.
Sources and scope
All sources are GitHub's own changelog and documentation pages.
- Upcoming deprecation of selected models in December, changelog, November 13, 2026
- About default availability of features and models, docs, accessed November 16, 2026
- Supported AI models, docs, accessed November 16, 2026
- Models and pricing, docs, accessed November 16, 2026
- Launch notes in the changelog: Anthropic's 5.2 release (November 9), Google's November 10 release and Google's Hydrogen (November 13), 2026
These sources cover Copilot availability, policy behavior and list prices. They don't say whether any alternative matches the retiring model's output on a given workload, or when a Kimi K3 replacement will be named. BaristaLabs did not run the retiring or replacement models while writing this.
Copilot model retirement
Check your Copilot model policies before December 14
Share your enterprise or organization model settings and the models your teams pick most. BaristaLabs can help find the swaps your default policy won't make, settle the Fable 5.2 retention question, and estimate what the replacements cost at current list prices.
A screenshot or export of your model settings is enough to start.
Turn this idea into a pilot
Which workflow should go first?
Use the readiness check to compare impact, effort, risk, owner, and next step before requesting a review.
- 3-5 minutes
- Deterministic score
- No sensitive data
Practical AI Workflow Notes
Want more practical AI operations ideas?
Get short notes on applying AI inside real small-business workflows — from document handling and customer follow-up to internal reporting, compliance, and automation guardrails.