On this page · 12 sections
- What GitHub announced on 13 November
- Three months of churn in context
- The admin fix most teams will miss: Fable 5.2
- Kimi K3 has no named successor
- What the replacements cost
- A worked example for one team
- Admin checklist before 14 December
- Questions to settle with your team
- India-specific considerations
- Dates to put in the calendar
- How eCorpIT can help
- References
Summary. On 14 December 2026 GitHub will remove six models from every Copilot experience, including Chat, inline edits, ask and agent modes and code completions: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. The suggested replacements are Claude Fable 5.2, Claude Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon and Grok 4.7; Kimi K3's successor is still to be announced. GitHub says the alternatives switch on automatically for Copilot Business and Enterprise unless an admin has turned off the global default, but its own documentation keeps the Fable family and open-weight models such as Kimi K3 outside default enablement. Any organisation that relies on Fable 5 must enable Fable 5.2 by hand and accept Anthropic's data retention to do it. Prices move both ways: Haiku 5.5 costs a tenth of Haiku 4.5 per token, Fable 5.2 is 20% cheaper than Fable 5, Gemini 4 Hydrogen costs twice Gemini 3.8 Flash's promotional rate, and Gemini 4 Carbon doubles to $4 input and $20 output per million tokens from 1 December. Individual subscribers on annual plans keep Haiku 4.5. This is the second Copilot retirement batch in a month, after Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra went on 16 November.
About this article. This is for GitHub enterprise and organisation owners, platform and DevOps leads, and engineering managers who pay for Copilot Business or Copilot Enterprise. On 21 November 2026 we checked every model name, date, plan, policy rule and price against GitHub's changelog posts, the Copilot documentation on supported models, default availability, model policies and pricing, and the European Central Bank's reference rates. The cost example uses stated assumptions, not measured usage. GitHub changes model line-ups and promotional prices often, so re-check before you budget; no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What GitHub announced on 13 November
GitHub's changelog notice of 13 November 2026 lists six models with a single deprecation date of 14 December 2026, applied across all Copilot experiences. It also says no admin action is needed to remove the retiring models: they drop out of the model picker on the day. What admins do need to check is whether the replacements will be there for users.
The table below combines the deprecation notice, the launch posts for each replacement and the default availability rules.
| Retiring model | Suggested replacement | Replacement arrived in Copilot | Default enablement applies? |
|---|---|---|---|
| Claude Fable 5 | Claude Fable 5.2 | 9 November 2026 | No, policy off until enabled |
| Claude Haiku 4.5 | Claude Haiku 5.5 | 7 October 2026 | Yes |
| Gemini 3.8 Flash | Gemini 4 Hydrogen | 13 November 2026 | Yes |
| Gemini 4 Argon | Gemini 4 Carbon | 10 November 2026 | Yes |
| Grok 4.6 | Grok 4.7 | Listed as GA in the docs | Yes |
| Kimi K3 | To be announced | Not applicable | Kimi K3 itself was excluded |
We did not find a dated Copilot launch post for Grok 4.7; the supported models page lists it as generally available alongside Grok 4.6.
One exception applies to individuals. Claude Haiku 4.5 stays available to individual Copilot subscribers on annual plans, so the Haiku change does not touch them. GitHub made the same carve-out for Claude Opus 4.8 in November, which the docs say remains open to annual individual subscribers but not to monthly ones.
Three months of churn in context
The December batch is the latest step in a fast cycle. GitHub announced on 15 October that Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra would go on 16 November, and confirmed the removal that day, pointing users to Claude Opus 5.5 and to GPT-6.1 Sol or GPT-6 Luna. The 16 November notice adds a detail worth knowing: Claude Opus 5 left the model picker but remains available as a fallback when another model refuses a request at runtime, under existing authorised access.
Before that, the documentation's retirement history records Claude Opus 4.7, Gemini 3.5 Flash, Gemini 3.6 Flash and Kimi K2.7 Code going on 2 October, and Gemini 3.7 Flash, GPT-5 mini, GPT-5.4, GPT-5.4 mini and GPT-5.5 going on 19 October. Our map of the six models retired on 1 September shows how quickly suggested replacements have themselves been retired.
Gemini 4 Argon is the clearest case. It arrived in Copilot on 13 October as Google's newest frontier model. It will be gone two months later, replaced by Gemini 4 Carbon, which GitHub added on 10 November. Gemini 3.8 Flash makes way for Gemini 4 Hydrogen, which GitHub describes as Google's latest Flash model with better instruction following than 3.8 Flash.
Another retirement sits just after this batch. GitHub will deprecate MAI-Code-1.1-Flash on 18 December 2026 in favour of the latest MAI Code Flash model, following the launch of MAI Code 2 Flash, and its notice says Enterprise admins may need to enable the alternative through model policies. That is a separate notice with a separate date, so track it on its own line.
On 21 November the retirement history table on the supported models page still listed 16 November as its latest entry and did not yet show the December models. Do not treat that table as your only alert: GitHub's own docs point admins to the changelog for new model announcements.
The admin fix most teams will miss: Fable 5.2
The deprecation notice says that under default model enablement the suggested alternatives switch on automatically for Copilot Business and Enterprise, unless an admin turned off the global default or disabled the model explicitly. For four of the five named replacements that holds. For Claude Fable 5.2 it does not.
GitHub's default availability page lists the models that sit outside the default policy and stay disabled whatever the setting: pre-GA models, open-weight models such as DeepSeek and Kimi K3, and models not covered by GitHub's data retention agreement, which it names as Claude Fable 5, 5.1 and 5.2. The Fable 5.2 launch post says the same thing in operational terms: Business and Enterprise admins must enable the Fable 5.2 policy, which is off by default, and leaving it off keeps the model unavailable.
The reason is data handling. Fable 5.2 needs data retention by default so Anthropic can run safety classifiers that detect harmful use; Anthropic retains prompts and outputs for that purpose and does not use them to train its models. Turning the policy on counts as acknowledging that requirement. Every other Claude model in Copilot, apart from Fable 5, 5.1 and 5.2, continues under zero data retention.
There is a time limit on the alternative. Eligible enterprises can use the Fable models with zero data retention until the end of 2026 under a time-bound exemption while Anthropic rolls out Enterprise Frontier Safeguards; after that, continued use needs those safeguards. Eligibility goes through the GitHub account team, not GitHub Support, and approval does not enable the model: an admin still has to switch the policy on. The docs add that customers who use a Fable model with zero data retention agree to use it only for internal operations and not to make its endpoints or outputs available externally.
So a team that runs Fable 5 today faces two decisions in December, two and a half weeks apart:
- Before 14 December, enable Fable 5.2 and accept data retention, or move users to a different model before Fable 5 disappears.
- Before 31 December, if you rely on the zero data retention exemption, find out from your account team what Enterprise Frontier Safeguards will require from January.
If your security or legal team signed off Fable 5 on the strength of zero data retention, Fable 5.2 needs a fresh sign-off on the same terms. If nobody signs it off, users who chose Fable 5 will find no Fable model in the picker on 14 December.
Kimi K3 has no named successor
Kimi K3 is the only model in the batch without a replacement; the notice says only that one will be announced. It was never on by default either, because GitHub treats open-weight models as outside the default enablement policy. Any team that uses it switched it on deliberately, usually for a reason such as price, behaviour on a codebase, or a preference for open weights.
GitHub's earlier pattern helps here. When Kimi K2.7 Code retired on 2 October, the suggested alternative in the docs was simply the latest Kimi model. Our guide to the Kimi K2.7 open-weight enable decision was written before that model retired, but the review it describes still applies to open-weight models. Whatever replaces Kimi K3 will most likely need the same explicit enablement, so plan the review now rather than in the week of 14 December. Until a successor is named, pick an interim model for the people who rely on Kimi K3 and tell them before the date.
What the replacements cost
Copilot bills by tokens converted into GitHub AI Credits, where one credit equals one US cent; Business and Enterprise get per-user allowances pooled at the billing entity, and usage beyond them is charged at the per-token rates. So a replacement with a different price changes your bill even if usage stays flat. Each replacement launch post we read says the model is billed at provider list pricing under usage-based billing.
All prices in the table come from GitHub's models and pricing page as checked on 21 November 2026.
| Model pair | Old input / output per 1M tokens | New input / output per 1M tokens | Change |
|---|---|---|---|
| Fable 5 to Fable 5.2 | $10.00 / $50.00 | $8.00 / $40.00 | 20% cheaper |
| Haiku 4.5 to Haiku 5.5 | $1.00 / $5.00 | $0.10 / $0.50 up to 100K input | 90% cheaper |
| Gemini 3.8 Flash to Gemini 4 Hydrogen | $0.75 / $3.75 promotional | $1.50 / $7.50 | Twice the price |
| Gemini 4 Argon to Gemini 4 Carbon | $2.00 / $10.00 promotional | $4.00 / $20.00 from 1 December | Twice the price |
| Grok 4.6 to Grok 4.7 | $2.00 / $6.00 up to 200K input | $2.00 / $6.00 up to 200K input | No change |
| Kimi K3 to successor | $3.00 / $15.00 | Not announced | Unknown |
Three caveats from the same page. Haiku 5.5 has a long-context tier above 100,000 input tokens at $0.50 input and $2.50 output, still half Haiku 4.5's rate. Gemini 3.8 Flash's promotional price runs to 31 December, but the model leaves on 14 December, so the cheap rate ends early for Copilot users. And Gemini 4 Argon's promotion ends on 1 December with no post-promotion price published, while Carbon, which carries the same promotional rate until then, moves to $4 and $20 from that day.
Anthropic models also carry a cache write cost, and Fable 5.2's cached input price of $0.20 per million tokens is lower than Fable 5's $1.00. For agent sessions that reuse large contexts, that difference can matter as much as the headline rate.
A worked example for one team
The assumptions below are ours, chosen for illustration. Your usage will differ, and cache reads and writes are left out to keep the arithmetic readable.
Team 40 developers on Copilot Business
Tokens per developer a month 30M input, 3M output on the one model
Gemini 3.8 Flash (promo) 30 x $0.75 + 3 x $3.75 = $33.75 per dev
Gemini 4 Hydrogen 30 x $1.50 + 3 x $7.50 = $67.50 per dev
Team change 40 x ($67.50 - $33.75) = +$1,350 a month
Claude Haiku 4.5 30 x $1.00 + 3 x $5.00 = $45.00 per dev
Claude Haiku 5.5 30 x $0.10 + 3 x $0.50 = $4.50 per dev
Team change 40 x ($4.50 - $45.00) = -$1,620 a month
Claude Fable 5 30 x $10.00 + 3 x $50.00 = $450.00 per dev
Claude Fable 5.2 30 x $8.00 + 3 x $40.00 = $360.00 per dev
Team change 40 x ($360.00 - $450.00) = -$3,600 a month
The Gemini Flash move adds $1,350 a month for this team, or 135,000 AI credits at one cent each. At the European Central Bank's 20 November 2026 reference rates, crossed to ₹92.68 per US dollar and rounded to two decimals, that is about ₹1,25,118 a month (1,350 x 92.68). The Fable move saves $3,600, about ₹3,33,648, but only if the team turns Fable 5.2 on. Haiku 5.5's figure assumes requests stay under the 100,000-token threshold; long-context requests would cost more but still less than Haiku 4.5.
The point of the exercise is not the totals. A model swap GitHub frames as a like-for-like replacement can double one line of your Copilot spend while cutting another, and the direction depends on which models your developers use. If you have not yet set budgets on the shared credit pool, our guide to Copilot AI credit pools and cost controls covers cost centres, per-user caps and chargeback.
Admin checklist before 14 December
These steps follow GitHub's documentation for enterprise and organisation owners on Business and Enterprise plans.
- Check the default policy. In the enterprise, open AI controls, then Copilot in the sidebar, then Configure models. Models that follow the default carry the Delegate to Default Policy label; if the enterprise-wide default availability policy is disabled, none of the replacements will switch on by themselves.
- Look for explicit disables. A replacement you disabled earlier, for example Gemini 4 Carbon during a security review, stays disabled. The deprecation notice is clear that auto-enablement does not override an explicit setting.
- Decide on Fable 5.2. It needs an explicit policy and the data retention acknowledgement described above, whatever your default.
- Plan for Kimi K3 users. Choose an interim model now, and expect any successor to need explicit enablement as an open-weight model.
- Check organisation-level settings. Where the enterprise delegates a model to organisations, each organisation owner sets it under Settings, Copilot, Models. A delegated model that an organisation has not configured follows that organisation's own default availability setting.
- Review Copilot Fusion models. The enterprise page has a separate list of models available to Fusion, and GitHub warns that disabling a model Fusion Max needs makes Fusion Max unavailable.
- Tell users and update scripts. Pinned model names in agent configurations, CLI scripts and team instructions will stop working on 14 December. GitHub asks customers to update workflows and integrations before the date.
- Update IDEs. GitHub's docs list minimum IDE and extension versions for some recent models and say the list is tentative, so keep Visual Studio, VS Code, JetBrains and Xcode plugins current before the switch.
Audit model policies with the API
Large enterprises with many organisations can check the state of every model in one pass. Since 19 October GitHub offers generally available GET and PATCH REST endpoints for model policies at the enterprise, enterprise team, agent app and organisation levels. A minimal read with the GitHub CLI looks like this:
# Enterprise-level model policies
gh api /enterprises/YOUR-ENTERPRISE/copilot/model-policies
# One organisation's model policies
gh api /orgs/YOUR-ORG/copilot/model-policies
Loop the organisation call across your organisations and look for the five replacement models and Kimi K3. We have not reproduced the response schema here; check GitHub's REST documentation before you script a PATCH. Teams that already manage Copilot through files can compare this with our guide to managed settings and team model policy.
What Auto users will see
Developers who leave the picker on Auto are affected too. Auto chooses from supported models subject to your policies and plan, and GitHub says the available models may change over time. Claude Haiku 4.5 and Gemini 3.8 Flash both appear in the list of models Auto and Copilot Fusion can choose, so after 14 December Auto will route to other models. For Business and Enterprise, GitHub keeps GPT-5.3-Codex as both the base model, used when no other model is enabled, and the long-term support model, which it commits to support for one year from its 18 March 2026 designation. If your policies leave nothing else enabled, that is what users fall back to.
Questions to settle with your team
- Who owns model approval? Two batches four weeks apart, plus a third retirement on 18 December, means model changes are now routine operations, not one-off events. Name an owner who reads the Copilot changelog weekly.
- Is the default policy right for you? Default enablement saves admin work but adds new models without review. Some enterprises keep it on and disable specific models; others turn it off only in organisations with stricter compliance needs, which GitHub's docs describe as an option.
- Do you need Fable at all? The Fable family is the only Claude line in Copilot that retains data by default. If the work can run on Claude Opus 5.5 or another model under zero data retention, the simplest fix is to move users there.
- Are budgets set on the shared pool? With Hydrogen costing twice what it replaces and Carbon doubling on 1 December, a credit budget is the control that stops a quiet price rise becoming a surprise invoice.
India-specific considerations
Indian engineering teams that deliver for US and European clients often run Copilot inside the client's enterprise account, so the model policy decisions above belong to the client's admins, not the vendor's. Raise the Fable 5.2 data retention question with the client's security team in writing before 14 December, because enabling the policy is the client's acknowledgement, not yours. Where your own organisation runs Copilot Business for Indian developers, budget in both currencies: at ₹92.68 per dollar, each extra $1,000 a month of model spend is about ₹92,680. For work that involves personal data of Indian users, check the retention change against your obligations under the Digital Personal Data Protection Act, 2023 with counsel; this article is not legal advice.
Dates to put in the calendar
- 1 December 2026: Gemini 4 Carbon moves to $4 input and $20 output per million tokens, and Gemini 4 Argon's promotional price ends.
- 14 December 2026: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 leave Copilot.
- 18 December 2026: MAI-Code-1.1-Flash leaves Copilot.
- 31 December 2026: The zero data retention exemption for Fable models ends; continued use needs Enterprise Frontier Safeguards.
- Any day: GitHub names a Kimi K3 successor, and keeps adding models such as Gemini 4 Neon, which arrived on 18 November.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our DevOps services team can audit your Copilot model policies across enterprises and organisations, script the checks against GitHub's model policy API, and set credit budgets before the December changes reach your bill. Talk to us at /contact-us/.
Last updated: 21 November 2026.