GitHub Copilot Model Deprecations: Dec 14 Migration and BYOK

GitHub will remove six hosted Copilot models on December 14, 2026. Replace Claude Fable 5, Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, and Grok 4.6; plan separately for Kimi K3. Check annual-plan exceptions, enable suggested models under Business and Enterprise policies, update hard-coded references, and use supported BYOK or external APIs when you need older models.

Contents
GitHub Copilot Model Deprecations: Dec 14 Migration and BYOK

GitHub Copilot will retire six hosted models on December 14, 2026, affecting developers who select those models manually or reference them in integrations. Switch to GitHub’s suggested replacements before the deadline, but check administrator policies because some replacements require explicit approval. If you need Kimi K3, Gemini 4 Argon, or Claude Fable 5 afterward, a compatible API endpoint offers another route without restoring the retired GitHub-hosted models.

GitHub confirmed the date and replacement list in its November 13 deprecation announcement. As of November 16, 2026, the important distinction is between a model disappearing from Copilot’s hosted catalog and that model becoming unavailable through every API provider.

Which GitHub Copilot models retire on December 14?

Six models are scheduled for retirement across Copilot Chat, inline edits, ask and agent modes, and code completions.

Retiring Copilot modelRetirement dateGitHub’s suggested alternative
Claude Fable 5December 14, 2026Claude Fable 5.2
Claude Haiku 4.5December 14, 2026*Claude Haiku 5.5
Gemini 3.8 FlashDecember 14, 2026Gemini 4 Hydrogen
Gemini 4 ArgonDecember 14, 2026Gemini 4 Carbon
Grok 4.6December 14, 2026Grok 4.7
Kimi K3December 14, 2026To be announced

GitHub explicitly exempts individual Copilot subscribers on annual plans from the Claude Haiku 4.5 retirement.

The Haiku exception applies to individual annual subscribers. GitHub does not extend it to monthly individual plans or Business and Enterprise seats in this announcement.

For Kimi K3, “To be announced” means GitHub has not named a replacement in the November 13 announcement. Do not interpret Kimi K3.1 Pro or another Kimi model as an officially designated Copilot successor.

The retirement covers GitHub-hosted model access in Copilot. It does not announce the shutdown of these models across independent provider APIs.

GitHub says users do not need to remove retired models manually. However, saved selections, integration identifiers, and team configurations still need migration before the cutoff.

How should individual developers switch models?

If you currently select a retiring model in Copilot Chat, choose GitHub’s suggested replacement in the model picker and test your normal coding task.

For Visual Studio Code:

  1. Open Copilot Chat.
  2. Open the model selector at the bottom of the Chat view.
  3. Select the suggested replacement, if your plan and organization allow it.
  4. Start a conversation and repeat a representative task, such as reading a repository file or explaining a failing test.
  5. Confirm the model responds and that your expected tools work.

GitHub documents the picker in Changing the AI model for Copilot Chat.

Changing the Chat model does not automatically change the model used for inline code suggestions. Check that feature separately if it is part of your workflow.

Choose by task and plan availability

GitHub’s suggested replacements are migration targets. Their capabilities, context limits, tool behavior, and costs are not guaranteed to match the retiring models.

Gemini 3.8 Flash users: Start with Gemini 4 Hydrogen for fast coding assistance and multistep tool use.

GitHub announced Gemini 4 Hydrogen as generally available on November 13. GitHub describes it as designed for fast agentic coding, tool use, and multistep workflows, with improved instruction following over Gemini 3.8 Flash.

Hydrogen is available to Copilot Pro, Pro+, Max, Business, and Enterprise users, subject to gradual rollout. The announcement lists VS Code, Visual Studio, Copilot CLI, GitHub’s coding agent, the Copilot app, github.com, mobile clients, JetBrains, Xcode, Eclipse, and Copilot SDK.

If Hydrogen is missing, check your plan, rollout status, and organization policy. An absent picker entry during rollout does not establish permanent unavailability.

For direct API changes outside Copilot, see the Gemini 4 Hydrogen migration and pricing guide.

Gemini 4 Argon users: Start with Gemini 4 Carbon for complex coding and reasoning tasks, including security-focused development.

GitHub’s November 10 Carbon announcement describes those workloads explicitly. Carbon is generally available in Copilot for Pro+, Max, Business, and Enterprise, with gradual rollout.

A Copilot Pro user should not assume Carbon is available merely because GitHub recommends it as an Argon replacement. Check plan eligibility before building a migration around it.

For applications that call Google’s API directly, consult the Gemini 4 Carbon API migration guide for model parameters and billing.

Claude Fable 5 users: Start with Claude Fable 5.2 for complex, multistep coding across large codebases.

GitHub’s November 9 release identifies that workload and reports improved task completion over Fable 5.1 in early testing. That is GitHub’s reported evaluation, not a guaranteed result for your repository.

Fable 5.2 is generally available to Pro+, Max, Business, and Enterprise users. Business and Enterprise administrators must explicitly enable it because its policy is off by default.

For independent coding-task comparisons, see the Coding Agent Index analysis of Fable 5.2 and Opus 5.5.

Claude Haiku 4.5 and Grok 4.6 users: Start with GitHub’s listed successors, Haiku 5.5 and Grok 4.7. Verify each model in your picker and run a representative task before changing defaults.

Kimi K3 users: GitHub has not announced a Copilot successor. If you must stay within GitHub-hosted Copilot, compare currently available models on your own tasks.

Kimi K3.1 Pro is a separate API candidate, not GitHub’s designated replacement. The Kimi K3.1 Pro coding guide covers that external option.

Check billing before changing defaults

Model availability and usage costs are separate decisions.

GitHub states that Gemini 4 Hydrogen, Gemini 4 Carbon, and Claude Fable 5.2 use provider list pricing under Copilot’s usage-based billing.

The Copilot models and pricing reference says interactions consume input, output, and cached tokens. Each category is priced according to the model, and the result is converted into AI credits at $0.01 USD per credit.

Included AI-credit allowances vary by plan. Additional usage beyond those allowances is billed at the documented per-token rates.

Before changing a team default, compare the same coding task, token usage, tool calls, and total charges under the relevant plans. Do not compare subscription allowances with direct API token prices as though they were equivalent.

What must Copilot Business and Enterprise administrators enable?

Under GitHub’s default model-enablement policy, suggested replacement models are generally enabled automatically for Copilot Business and Enterprise.

That automatic behavior does not apply when the global default is disabled, a model is explicitly restricted, or the model is excluded from default enablement.

Claude Fable 5.2 is an exception. GitHub’s November 9 announcement says its administrator policy is off by default and must be enabled explicitly.

Organization administrators

If your organization controls its model policies:

  1. On GitHub, open Organizations and select your organization.
  2. Open Settings.
  3. Under Code, planning, and automation, select Copilot.
  4. Open Models.
  5. Find each replacement and inspect its current setting.
  6. Choose Enabled for models the organization permits.

These steps follow GitHub’s organization model-availability instructions.

The page also exposes Default availability for released models. A model marked Delegate to Default Policy inherits the default when that model is eligible.

If an enterprise owner has enforced a model setting, organization owners cannot override it. Ask the enterprise administrator to review that restriction.

Enterprise administrators

Enterprise owners control the baseline centrally:

  1. Open the enterprise account on GitHub.
  2. Select AI controls.
  3. Open Copilot.
  4. Select Configure models.
  5. Enable approved replacements for everyone, or delegate the decision to organizations or enterprise teams.
  6. Verify that the intended organizations or teams have access.

GitHub documents these controls in Managing availability of models in your enterprise.

If your enterprise opts into the enterprise-teams model-access preview, organization-level model settings are deactivated. Configure delegated access through enterprise teams instead.

Why Fable 5.2 needs separate approval

GitHub’s Fable 5.2 release notice says Anthropic retains data, including prompts and outputs, to operate safety classifiers when Fable 5.2 is used.

GitHub says the retained data is not used to train Anthropic’s models. Enabling the Fable 5.2 policy acknowledges the retention requirement.

Administrators should review this requirement before enabling the model for repositories containing sensitive code or data.

GitHub also documents a time-bound zero-data-retention exemption for eligible enterprise customers using Fable 5, Fable 5.1, or Fable 5.2. The exemption runs through the end of calendar year 2026 while Anthropic introduces Enterprise Frontier Safeguards (EFS).

After that period, continued use under the announced arrangement requires EFS. GitHub says EFS will provide automated safety monitoring and additional retention options for eligible customers.

Contact the GitHub account team to determine eligibility. GitHub Support cannot approve access, and exemption approval does not automatically enable the model policy.

If your organization cannot accept the standard retention terms and has no approved exemption, choose another permitted model.

How to verify that the policy works

Ask an affected user to reopen Copilot Chat after the administrator changes the policy.

The enabled replacement should appear in the model selector in a supported experience. Select it and complete a normal coding prompt.

If it remains missing, check the enterprise baseline, organization delegation, explicit restrictions, plan eligibility, and rollout status.

A visible model and a successful task provide a stronger migration check than the administrator’s saved policy alone.

How do you update integrations that hard-code retiring models?

A model-picker change does not update source code, scripts, configuration files, or integrations that specify a retiring model.

Search for both display names and API-style identifiers. Exact identifiers differ between GitHub clients and independent model providers.

From a Bash terminal in your repository, use ripgrep to find likely references:

rg -n -i 'Claude Fable 5|Claude Haiku 4\.5|Gemini 3\.8 Flash|Gemini 4 Argon|Grok 4\.6|Kimi K3|claude-fable-5|claude-haiku-4-5|gemini-3\.8-flash|gemini-4-argon|grok-4\.6|kimi-k3' .

This is an inventory search, not an automatic migration. It can miss alternate aliases, values outside the repository, and files excluded by .gitignore.

Review matches in:

  • Copilot SDK initialization and session configuration.
  • Agent configuration and model-routing rules.
  • Repository automation that invokes Copilot.
  • CI configuration and environment variables that select a model.
  • Internal tools that call models by stored identifiers.

Replace each pinned model setting with an identifier supported by that client. Do not assume the display name Claude Fable 5.2 is a valid API identifier everywhere.

Keep provider endpoints and authentication settings unchanged unless the new model’s documented integration requires an update.

For agent integrations, verify the complete operation:

  1. Start a fresh session using the replacement.
  2. Ask the agent to read a known repository file.
  3. Request a small, reversible code change or tool call.
  4. Inspect the result and run the relevant tests.
  5. Check usage records for the selected model and unexpected errors.

A successful text response alone does not verify file editing or tool calling.

If the replacement is missing, inspect access policies first. If the integration rejects its identifier, consult the client’s supported-model documentation.

Finish the inventory before December 14. GitHub’s automatic removal of retired models will not rewrite your configuration.

Can you keep using Kimi K3, Gemini 4 Argon, or Fable 5?

You can continue through an independent API provider while the model remains available and usable there.

The Copilot retirement announcement concerns GitHub-hosted model access. It does not announce a corresponding shutdown across other API providers.

As of November 16, 2026, BetterToken’s structured model catalog lists these identifiers:

ModelBetterToken API model ID
Kimi K3kimi-k3
Gemini 4 Argongemini-4-argon
Claude Fable 5claude-fable-5
Claude Haiku 4.5claude-haiku-4-5-20251001

These are BetterToken API IDs, not GitHub-hosted Copilot model identifiers.

BetterToken provides separate pay-as-you-go API access. It is not a GitHub Copilot subscription or an official model-vendor subscription.

A catalog listing does not guarantee a successful request or permanent availability. Check the currently offered model and pricing before relying on an older model in production.

Option 1: Continue with Cline or OpenCode

For developers who want a configurable coding assistant, Cline supports an OpenAI-compatible provider.

Follow the BetterToken Cline setup guide:

  1. Install Cline in VS Code and create your own BetterToken API key in the console.
  2. Open Cline and choose Use your own API key.
  3. Set API Provider to OpenAI Compatible.
  4. Set Base URL to https://www.bettertoken.ai/v1.
  5. Enter your key in Cline’s credential field and set Model to kimi-k3, if it remains available.
  6. Save the configuration and reload Cline if needed.
  7. Use Verify, then send a short test prompt.

Confirm that the response arrives without an authentication or model error. Next, request a read-only operation on a known project file.

Configure the selected model’s confirmed token limits. Enable tool, image, or Computer Use capabilities only when the model and endpoint support them.

OpenCode offers another documented route. Its OpenAI-compatible configuration uses @ai-sdk/openai-compatible, the Base URL https://www.bettertoken.ai/v1, and a provider ID such as bettertoken.

For OpenCode 1, save the provider configuration in ~/.config/opencode/opencode.json or a project-level opencode.json. Use /connect to save credentials locally rather than embedding a real API key in JSON.

The documented Kimi K3 selection is bettertoken/kimi-k3. See the OpenCode Kimi K3 setup instructions for the complete configuration.

Claude Code uses a different protocol. Its BetterToken integration guide specifies the Anthropic-compatible Base URL https://bettertoken.ai, without /v1.

Do not reuse Cline’s OpenAI-compatible Base URL as the Claude Code Base URL.

Option 2: Verify the model with a direct API request

Before investigating a tool-specific failure, confirm that the model can answer through the provider’s API.

BetterToken documents an OpenAI-compatible Chat Completions endpoint. The following Bash example requests a short response from kimi-k3.

read -rs BETTERTOKEN_API_KEY && export BETTERTOKEN_API_KEY

curl -sS -i 'https://www.bettertoken.ai/v1/chat/completions' \
  -H "Authorization: Bearer ${BETTERTOKEN_API_KEY}" \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "kimi-k3",
    "messages": [
      {
        "role": "user",
        "content": "Reply with OK."
      }
    ]
  }'

Paste your API key at the silent prompt and press Enter. The command references the exported environment variable without placing the key directly in the request text.

A successful request returns HTTP 200 with model output in choices[0].message.content.

Another model ID from the table can use this request format if its current catalog entry supports OpenAI-compatible Chat Completions.

If the call fails, use the documented checks:

ResultFirst check
400Request JSON, model parameters, and message structure
401API key and Bearer authentication
402Available account balance
404Direct request URL; use the complete /v1/chat/completions path
429Rate limits, concurrency, or upstream load
Model not foundExact Model ID from the current catalog

The BetterToken Chat Completions reference distinguishes direct HTTP URLs from SDK Base URLs. For a direct request, use https://www.bettertoken.ai/v1/chat/completions. For an SDK that appends its own path, use https://www.bettertoken.ai/v1.

A successful text request does not prove that streaming, tool calls, or the full model context will work. Verify those capabilities in the target coding client.

Does GitHub Copilot BYOK support custom API endpoints?

Yes, in supported clients. GitHub’s Bring your own key documentation describes user-configured models that connect directly to compatible cloud or local providers.

As of November 17, 2026, GitHub’s BYOK-specific client table lists VS Code, JetBrains, Xcode, Copilot CLI, the GitHub Copilot app, and Copilot SDK. These are supported client categories, not a promise that every client accepts every protocol or model.

Their setup methods and supported capabilities differ. Adding a custom model does not restore a retired model to GitHub’s hosted catalog.

Configure a custom endpoint in VS Code

VS Code’s Custom Endpoint provider supports Chat Completions, Responses, and Anthropic Messages APIs. For a BetterToken Chat Completions connection, use the complete request URL rather than an assumed provider-specific shortcut.

You need a compatible VS Code installation, the model ID, and your own provider API key.

Follow the documented VS Code Custom Endpoint setup:

  1. Run Chat: Manage Language Models from the Command Palette.
  2. Select Add Models, then Custom Endpoint.
  3. Enter a group name, such as BetterToken.
  4. Enter the display name and API key when prompted. Keep the secret reference generated by VS Code.
  5. Select Chat Completions as the API type.
  6. Open the generated chatLanguageModels.json and configure the model properties.
  7. Save the file, then select the custom model in the Chat model picker.

The following is an illustrative chatLanguageModels.json configuration for kimi-k3. Keep the actual apiKey reference generated by VS Code instead of copying the example secret-reference identifier.

[
  {
    "name": "BetterToken",
    "vendor": "customendpoint",
    "apiKey": "${input:chat.lm.secret.xxxxx}",
    "apiType": "chat-completions",
    "models": [
      {
        "id": "kimi-k3",
        "name": "Kimi K3",
        "url": "https://www.bettertoken.ai/v1/chat/completions",
        "toolCalling": false,
        "vision": false,
        "maxInputTokens": 8192,
        "maxOutputTokens": 2048
      }
    ]
  }
]

The token values in this example are deliberately small client-side settings for an initial text test. They are not Kimi K3’s published context or output limits. Replace them with values appropriate to the current model documentation and workload.

Keep the combined maxInputTokens and maxOutputTokens within the model’s actual context window.

The initial configuration disables tool calling and image input. Enable either capability only after confirming support from the model, endpoint, and client. Agent mode requires tool calling.

VS Code’s vendor value must be customendpoint. The model id is sent to the API, while name is shown in the picker.

The apiType value selects the request protocol. The url above is the full Chat Completions endpoint, not an SDK Base URL.

VS Code may append a resource path automatically when a shorter URL is supplied. The Custom Endpoint URL-resolution rules recommend the explicit API path to avoid ambiguity.

VS Code also supports secret-input references in apiKey. Its setup flow stores the entered credential locally and generates a reference such as ${input:chat.lm.secret...}. Do not replace that generated reference with the key itself.

After saving, reload the VS Code window if the model does not appear. Select kimi-k3 and send a short text prompt.

If the request succeeds, proceed to a separate tool-calling test before using it for repository edits.

What BYOK does not restore

VS Code documents BYOK for chat and utility tasks. Native inline suggestions, semantic search, and certain service-backed features still require GitHub account access.

For supported BYOK models in Copilot harness sessions, VS Code currently requires the experimental chat.agentHost.byokModels.enabled setting. Enabling it does not guarantee that every external model supports all agent functions.

Business and Enterprise administrators can also restrict user-configured BYOK through organization or enterprise policies.

For Copilot CLI, GitHub documents custom providers in ~/.copilot/providers.json. The openai provider type supports OpenAI-compatible services, while anthropic supports Anthropic-style providers.

However, GitHub warns that credentials stored in this CLI JSON file are plain text. Environment-variable strings are not automatically expanded. See Adding LLM models to Copilot CLI before using this route.

Enterprise-managed custom models

GitHub also supports enterprise-managed custom models in public preview.

Enterprise owners can add provider API keys in enterprise settings. They can also allow organization owners to configure models through the Enable custom models policy.

GitHub lists OpenAI-compatible providers among the supported categories. Custom models are available through Copilot Chat, Copilot CLI, and supported IDE experiences for eligible users on the enterprise’s Copilot Business or Enterprise plan.

Unlike user-configured BYOK, these centrally managed models are accessed through GitHub’s Copilot API. Users need a Copilot license and internet access but do not configure the provider key themselves.

The enterprise custom-model instructions describe how to add provider credentials and grant organization access.

The feature remains a public preview. Test model compatibility and the required Copilot surface before using it as an organization-wide retirement workaround.

What should be verified before December 14?

Before the retirement date, check the replacement model in the actual client where developers work.

Individual users should confirm the model appears and can complete a representative coding task. Claude Haiku 4.5 users who want the exception must confirm they hold an eligible individual annual subscription.

Administrators should verify policies with an affected user, paying particular attention to Fable 5.2’s explicit enablement and data-retention requirements.

Integration owners should remove hard-coded retiring identifiers and test real tool calls, not only text responses.

If Kimi K3 or another older model remains necessary, validate an independently available API route before the cutoff. Keep external API availability, BYOK configuration, and GitHub-hosted Copilot model access separate.

Ready to optimize your LLM workflow?

Join thousands of developers building faster, smarter, and more cost-effective AI applications with BetterToken.

Get Started for Free