Tools

GitHub Copilot drops three models, VS Code retires Local harness

On 16 November 2026 Copilot retired Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra, GitHub set out VS Code's three-month exit from its Local harness, and admins gained a plugin-only setting. What to migrate, and by when.

GitHub made three changes on 16 November 2026 that land on the same people: the engineers who choose models in Copilot and the administrators who govern it. It deprecated three models, Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. It announced that VS Code will remove the Local agent harness in February 2027. And it gave enterprise administrators a setting to force Copilot customizations to come from plugins or managed settings rather than from users’ and projects’ own files.

The model change needs the least work: GitHub says suggested replacements are enabled automatically for most Business and Enterprise customers, and nothing has to be removed. The harness change needs the most. Moving from the Local harness to the Copilot harness does not convert prompt files, hooks or every MCP server configuration, and VS Code’s own guide warns that the same model can behave differently after the switch. The customization setting is optional, but it is the first control that lets an enterprise distribute approved MCP servers and skills through a plugin library while blocking the rest.

This is a migration guide built from GitHub’s changelog posts, GitHub’s documentation and VS Code’s documentation, read on 17 and 18 November 2026. GenAI Brief has not tested the migration tooling; where documents disagree, we say so.

The dates in one place

DateChangeSource
15 Oct 2026Mid-November deprecations announcedGitHub changelog
28 Oct 2026Microsoft announces the Copilot harness as a shared runtime across its Copilot productsMicrosoft blog
11 Nov 2026VS Code 1.146 released; deprecates the Local harnessVS Code release notes
13 Nov 2026Six December deprecations announcedGitHub changelog
16 Nov 2026Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra deprecated; Local harness retirement and plugin-only customizations postedGitHub changelog
17 Nov 2026MAI-Code-1.1-Flash deprecation announced, with MAI Code 2 Flash launchGitHub changelog
14 Dec 2026Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, Kimi K3 deprecatedGitHub changelog
18 Dec 2026MAI-Code-1.1-Flash deprecatedGitHub changelog
February 2027Local harness and vscode.temporaryDefaultToLocalHarness removedGitHub changelog, VS Code docs

Compiled by GenAI Brief. GitHub gives the harness removal as a month, not a day.

The model deprecations

The 16 November post lists three models and their suggested alternatives. It covers “most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions)”. The 15 October notice had said “all” experiences; the change of wording is not explained.

Deprecated modelDateGitHub’s suggested alternativeExceptions and notes
Claude Opus 4.816 Nov 2026Claude Opus 5.5Still available to individual subscribers on annual plans
Claude Opus 516 Nov 2026Claude Opus 5.5Removed from the model picker; “remains available for runtime refusal fallback through existing authorized access”
GPT-5.6 Terra16 Nov 2026GPT-6.1 Sol or GPT-6 LunaGitHub’s docs list “Latest GPT Terra model” as the alternative

Source: GitHub changelog, 16 November 2026, and the “Supported AI models” documentation, read 18 November 2026.

Two details in that table need explaining.

The Terra replacement differs between GitHub’s own pages. The changelog suggests GPT-6.1 Sol or GPT-6 Luna. The model retirement history in GitHub’s documentation lists the suggested alternative for GPT-5.6 Terra as “Latest GPT Terra model”, and the same page lists GPT-6.1 Terra as generally available. A team that picked GPT-5.6 Terra for its particular cost and speed profile should test GPT-6.1 Terra alongside the two models the changelog names, rather than assume Sol or Luna is the closer fit.

“Runtime refusal fallback” is not defined in the posts or documents we read. The changelog says Claude Opus 5 is gone from the picker but still serves “runtime refusal fallback through existing authorized access”, and that “existing enterprise model policies continue to apply”. GitHub does not say which models fall back to Opus 5, or when. The phrase suggests the kind of fallback described in our Copilot Fusion coverage, where another model completes a request after the selected one refuses on safety grounds; that example came from a benchmark run of Anthropic’s Claude Code harness, not from Copilot, so this is our inference. Administrators who disabled Opus 5 by policy should assume that policy still governs whether it can act as a fallback, since GitHub says existing policies continue to apply.

Who has to do anything

For Copilot Business and Enterprise, the changelog’s instruction is conditional: “Under default model enablement, the suggested alternatives are automatically enabled for Copilot Enterprise and Copilot Business customers unless an administrator has turned off the global default or explicitly disabled the model.” If the global default is off, the alternatives have to be enabled through their model policies in Copilot settings. “No action is required to remove the deprecated models.”

The work therefore falls on three groups:

  1. Enterprises that turned off default enablement. Opus 5.5, and GPT-6.1 Sol or GPT-6 Luna, need explicit policy approval, or users lose their Opus and Terra options entirely.
  2. Teams with hard-coded model names. Scripts, custom agents and managed defaults that name a deprecated model need updating. The managed model key in enterprise-managed settings takes a specific model and version string, "auto" or "fusion", according to the managed settings reference.
  3. Individual subscribers on monthly plans who used Opus 4.8. The documentation’s plan table footnote says Opus 4.8 “is not available to subscribers on monthly plans”.

GitHub’s documentation also describes a backstop. A base model is the default Copilot uses “when no other models are enabled”, and an LTS model is one GitHub commits to supporting for a year. GitHub designated GPT-5.3-Codex as both on 18 March 2026, for Business and Enterprise only, according to its base and LTS page. That year runs to March 2027.

December brings a harder case

Four weeks later, on 14 December 2026, six more models go, according to a 13 November notice. A seventh, MAI-Code-1.1-Flash, follows on 18 December.

ModelDateSuggested alternativeEnabled automatically under default enablement?
Claude Fable 514 Dec 2026Claude Fable 5.2No. Fable models are excluded from default enablement
Claude Haiku 4.514 Dec 2026Claude Haiku 5.5Per the notice, yes, unless the global default is off; annual individual plans keep Haiku 4.5
Gemini 3.8 Flash14 Dec 2026Gemini 4 HydrogenPer the notice, yes, unless the global default is off
Gemini 4 Argon14 Dec 2026Gemini 4 CarbonPer the notice, yes, unless the global default is off
Grok 4.614 Dec 2026Grok 4.7Per the notice, yes, unless the global default is off
Kimi K314 Dec 2026“To be announced”Not applicable; open-weight models are excluded from default enablement
MAI-Code-1.1-Flash18 Dec 2026“Latest MAI Code Flash model”The notice says Enterprise admins “may need to enable access”

Source: GitHub changelog, 13 and 17 November 2026; “Models eligible for default enablement” in GitHub’s supported models documentation, read 18 November 2026.

The Fable row is the one that catches administrators out. GitHub’s documentation lists three groups of models that are disabled by default “regardless of your ‘Default availability’ policy setting”: pre-GA models, open-weight models (it names DeepSeek and Kimi K3), and “Models that are not covered by GitHub’s data retention agreement (Claude Fable 5, Claude Fable 5.1, Claude Fable 5.2)”. The December notice repeats the boilerplate about automatic enablement, but by GitHub’s own rules that does not apply to Fable 5.2.

The reason is data handling. When GitHub made Claude Fable 5 available in June 2026, it said the model “requires data retention”: Anthropic retains prompts and outputs for up to 30 days to operate safety classifiers, while other Claude models in Copilot run under zero data retention. “Enabling the Claude Fable 5 policy constitutes acknowledgement of this requirement.” An enterprise that approved Fable 5 has accepted that once; enabling Fable 5.2 is a separate policy decision on the same default data-retention terms, and should go through whoever approved the first. GitHub’s supported models page now also describes a time-bound zero-data-retention exemption for Fable models, through the end of 2026, for enterprises its account team approves.

The Local harness retirement

The harness is the software around the model: it assembles context, runs tools, handles approvals and keeps the conversation. In VS Code, the Local harness is the older, editor-specific implementation. The Copilot harness is built on the Copilot SDK and, in GitHub’s words, “shares its coding-agent runtime with Copilot CLI and the GitHub Copilot app”. Microsoft announced on 28 October that it is standardising its Copilot experiences on that harness.

The VS Code 1.146 release notes, dated 11 November 2026, set the terms. The Copilot harness has been the default for new sessions since VS Code 1.143. For three months the Local harness remains usable “while you move your workflows, but it’s functionally frozen”. After that it “will no longer be available”, and the vscode.temporaryDefaultToLocalHarness enterprise-managed setting, introduced in VS Code 1.142, goes with it. GitHub’s changelog adds that “keeping the exception configured won’t keep the Local harness available.”

GitHub is also clear about what the switch is not. “Choosing the Copilot harness doesn’t move your files to the cloud,” and agent commands still run on your machine. Selecting it “doesn’t enable sandboxing” either; that is configured separately.

What does not convert by itself

VS Code’s migration guide is the most important document in this change. It says the Copilot harness “reads customizations from Copilot folders and portable formats”, so anything stored in editor-specific locations has to move. A notice in chat flags workspaces that need migrations, and Chat: Open Customizations > Migrations lists them.

Existing customizationWhere it goes in the Copilot harnessAutomatic?What to recheck
Workspace .vscode/mcp.json.mcp.json at the workspace rootYes, for compatible entriesAuthentication, tools, enabled state, sandbox
Profile mcp.json~/.copilot/mcp-config.json (or $COPILOT_HOME/mcp-config.json)Yes, for compatible entriesAs above
Prompt files (*.prompt.md)Project or user skillsConversion availableDropped agent, model or tools properties
Agents and instructions in profile storageCopilot user foldersMove availableDiscovery in a new session
HooksSupported Copilot hook sources and scriptsNoEvent timing, inputs and outputs, failure paths
User-defined tool setsWorkflow-specific tool configurationNoRequired capabilities
Chat participantsA publisher replacement or a direct agent requestNoService access and context
Local harness memory filesCopilot Memory (cloud-backed by default)NoKeep required guidance in instructions or skills

Source: VS Code “Migrate customizations to the Copilot harness” and “Move from the Local to the Copilot harness”, read 18 November 2026.

The MCP details are where secrets and safety settings can quietly change:

  • Input prompts. VS Code’s format can prompt for an API key through ${input:api-key}. The Copilot format does not use that flow. Non-secret inputs migrate to environment-variable references; “password and command inputs require manual changes”. The guide is explicit: “Environment variables are not a secret store.”
  • Tool scope widens. A server with standard values “migrates automatically and adds tools: ["*"]”, so every tool the server exposes becomes available until someone narrows the list.
  • Disabled servers may come back. “Disabled state is not stored in .mcp.json or mcp-config.json.” User-level disablement can be lost, and other clients can treat the migrated server as enabled.
  • Sandbox settings do not transfer. Per-server and top-level sandbox rules from VS Code do not carry over; the Copilot harness uses session-level sandbox settings, and remote HTTP and SSE servers “are not locally sandboxed”.
  • Source entries are removed. Automatic MCP migration writes and verifies each destination entry, then removes the migrated source entry. Migrate with Agent, the agent-led alternative, keeps sources by default and uses Copilot credits.

Prompt files are the other common break. “Copilot harness sessions use agent skills for reusable workflows and do not load *.prompt.md files.” Prompt files still work in the Local harness until it is removed.

A worked migration for one developer

A hypothetical developer has three things: a workspace .vscode/mcp.json with a database server that prompts for a password, two prompt files the team uses for release notes, and a pre-commit hook defined in VS Code settings.

Following the documentation, the migration list would offer to move the MCP server but would put it under Needs manual review, because password inputs do not migrate. The developer replaces the input with a ${DB_PASSWORD} reference, provides the variable through the team’s approved mechanism on the machine where the agent runs, narrows tools from ["*"], and checks the server’s enabled state. The prompt files convert to two skills; if either set a model, that has to be re-added. The hook has no automatic conversion and has to be rewritten against the Copilot hook events, then tested for what happens when it fails. Only then is the source configuration retired. The guide’s advice to back up first, run a small known task before and after, and avoid leaving duplicate active copies applies at every step.

What administrators control

For organisations, VS Code’s transition guide recommends a pilot, then requiring the harness for a rollout group. The keys belong in enterprise-managed settings, not user or workspace settings:

{
  "vscode": {
    "temporaryAllowLocalHarness": false,
    "temporaryDefaultToLocalHarness": false
  }
}
SettingEffect, per VS Code’s documentation
vscode.temporaryAllowLocalHarness: falseProhibits the Local, Claude and Codex harnesses; takes precedence over the temporary Local default
vscode.temporaryDefaultToLocalHarness: trueKeeps Local available and stops automatic routing to Copilot; removed in February 2027
ChatEditorPreferCopilotHarness device policySelects Copilot for new editor chats without prohibiting other harnesses
Non-empty managed permissions.deny, or sandbox.enabled: trueFrom VS Code 1.143, require the Copilot harness unless the temporary Local default applies
Any Copilot enterprise-managed settingsDisable the Claude and Codex harness integrations in VS Code

Source: VS Code “Move from the Local to the Copilot harness”, “Choose an agent harness” and “Manage AI settings”, read 18 November 2026.

The last row deserves attention in any rollout plan. VS Code’s enterprise documentation says the Claude and Codex harnesses “are disabled while Copilot enterprise-managed settings apply”, including server-managed, MDM and file-based deployments, because “the Copilot runtime enforces these settings, but the Claude and Codex harnesses cannot enforce them.” Deploying a single managed setting, even one unrelated to harnesses, can therefore remove harnesses that some developers rely on.

The reason to move, from an administrator’s side, is enforcement. GitHub says the Copilot harness lets administrators “enforce managed runtime permissions and sandbox restrictions that the Local harness cannot enforce”. The transition guide warns that falling back to Local while keeping the same managed configuration “does not preserve those protections”, and asks administrators to “review and accept this gap before permitting fallback.” VS Code’s enterprise page makes the same point for hooks: “Local harness sessions do not load SDK Policy Hooks.”

Verification is part of the documented procedure: run Developer: Policy Diagnostics to see effective policy values and their sources, run /sandbox policy inside a Copilot session, and test one operation policy should allow and one it should block.

Plugin-only customizations

The third change gives administrators a way to make the migration stick. The strictPluginOnlyCustomization managed setting “blocks standalone customizations from user and project sources”, while “eligible plugins and managed settings remain allowed, subject to your other policies”. It can cover all five supported categories or selected ones: skills, custom agents, hooks, MCP servers and executable Copilot extensions. GitHub clarifies that extensions here “include executable canvas providers, not VS Code Marketplace extensions”, and gives the example of restricting MCP servers and extensions while leaving skills, agents and hooks open.

Plugin-only is not a marketplace allowlist. “Plugin-only does not mean enterprise-managed plugins only,” the changelog says. Without further controls, a user could still install an unvetted plugin from a public marketplace that bundles an MCP server. The pairing GitHub recommends uses the existing plugin keys in managed-settings.json, which lives in a .github-private repository at copilot/managed-settings.json for server-managed deployments:

KeyWhat it does
strictPluginOnlyCustomizationBlocks standalone user and project customizations in the chosen categories
strictKnownMarketplacesRestricts plugin installation to listed marketplaces and disables the built-in ones; an empty array is complete lockdown
extraKnownMarketplacesAdds marketplaces, for example an internal GitHub repository; members need read access to it
enabledPluginsRequires (true) or blocks (false) named plugins, keyed PLUGIN-NAME@MARKETPLACE-NAME
allowedMcpServers / deniedMcpServersAllowlist or denylist MCP servers by name, URL or exact command

Source: GitHub changelog, 16 November 2026, and GitHub’s managed settings reference and plugin marketplace guide, read 18 November 2026.

The setting is available on GitHub Enterprise Cloud “for members with Copilot Business or Copilot Enterprise licenses issued by the enterprise or one of its organizations.” Three documentation gaps are worth knowing before relying on it:

  • As read on 18 November 2026, GitHub’s managed settings reference does not yet list strictPluginOnlyCustomization in its supported keys or its client support table, so which clients enforce it, and from which versions, is not documented there.
  • VS Code has its own device policy, ChatStrictPluginOnlyCustomization, which its enterprise page describes as blocking “standalone user and workspace skills, agents, instructions, and MCP servers”. That list includes instructions and omits hooks and extensions, unlike the changelog’s five categories. VS Code pairs it with ChatAllowManagedHooksOnly for hooks.
  • The managed settings reference says enterprise-managed settings “can only require or block plugins”; optional project defaults go in a repository’s .github/copilot/settings.json.

A reasonable reading is that an enterprise wanting to restrict MCP servers to an approved library would set plugin-only for MCP servers and extensions, list its internal marketplace in strictKnownMarketplaces, require the plugins that carry approved servers with enabledPlugins, and keep allowedMcpServers as a second check. That is our inference from the documents, not a configuration GitHub publishes.

What GitHub has not said yet

Several facts an administrator would want are missing from the posts and pages we read.

An exact removal date for the Local harness. GitHub and VS Code give “February 2027” and “three months”. Neither names a day or a VS Code version. Plan for the start of February, not the end.

What “runtime refusal fallback” covers. GitHub says Claude Opus 5 still serves it, but not for which models, on which surfaces, or how long. If your audit requires a list of models that can process your code, ask GitHub whether Opus 5 belongs on it after 16 November.

Kimi K3’s replacement. The December notice says “To be announced”. Teams that enabled Kimi K3 deliberately, as an open-weight option, have less than four weeks and no named successor.

Client coverage for plugin-only customizations. The changelog does not list supported clients or minimum versions, and the managed settings reference does not yet include the key. VS Code’s device policy of a similar name lists a slightly different set of categories. Until GitHub publishes a support table, test enforcement in each client your developers use, including Copilot CLI and JetBrains IDEs.

How migrated customizations behave in other clients. VS Code’s migration guide says VS Code and Copilot CLI read the same Copilot customization files, but its transition guide adds that “supported formats can help you reuse customizations across clients; they don’t guarantee identical tools or automatic synchronization of personal files.” A skill that works in the editor is not proven to work in the CLI until someone runs it there.

What to do, in order

  1. By the end of November: check model policies. If your enterprise turned off default enablement, enable Claude Opus 5.5 and your chosen GPT replacement. Search scripts, custom agents and managed model values for the three deprecated models.
  2. Before 14 December: decide on Claude Fable 5.2. It will not switch on by itself. Route the decision through whoever accepted Fable 5’s data retention terms. Note that Kimi K3 has no named replacement yet.
  3. Before 18 December: enable MAI Code 2 Flash if teams use MAI-Code-1.1-Flash.
  4. Before the end of the year: inventory Local harness dependencies. Prompt files, hooks, MCP servers with input prompts, tool sets, chat participants. Pilot the Copilot harness on one team.
  5. During the pilot: verify, do not assume. Policy Diagnostics, /sandbox policy, and a test of one blocked operation. Narrow migrated tools: ["*"] entries and recheck disabled MCP servers.
  6. Before rolling out managed settings: warn Claude and Codex harness users that any enterprise-managed settings will disable those integrations in VS Code.
  7. Before February 2027: remove vscode.temporaryDefaultToLocalHarness exceptions. They stop working when the Local harness is removed.
  8. Optional: adopt plugin-only customizations for MCP servers first, paired with strictKnownMarketplaces, once clients confirm support.

None of these changes is large on its own. Together they move Copilot’s control surface from the editor to a shared runtime that GitHub can govern across clients, and they shift the default from users configuring their own tools to administrators distributing approved ones. Teams that treat the February deadline as the start date will be migrating hooks and secrets under time pressure; the documentation gives everything needed to start now.

Sources

  1. GitHub Copilot models deprecated on November 16, 2026 GitHub Changelog
  2. Upcoming deprecation of selected GitHub Copilot models in mid-November GitHub Changelog
  3. Upcoming deprecation of selected GitHub Copilot models in December GitHub Changelog
  4. Upcoming deprecation of MAI-Code-1.1-Flash in December GitHub Changelog
  5. VS Code retires the Local harness in February 2027 GitHub Changelog
  6. Control where GitHub Copilot customizations come from GitHub Changelog
  7. Supported AI models in GitHub Copilot GitHub Docs
  8. Base and long-term support (LTS) models GitHub Docs
  9. Enterprise managed settings GitHub Docs
  10. Managing plugin marketplaces and plugins with enterprise-managed settings GitHub Docs
  11. Getting started with enterprise-managed settings GitHub Docs
  12. Claude Fable 5 is generally available for GitHub Copilot GitHub Changelog
  13. Visual Studio Code 1.146 release notes Visual Studio Code
  14. Move from the Local to the Copilot harness Visual Studio Code documentation
  15. Migrate customizations to the Copilot harness Visual Studio Code documentation
  16. Choose an agent harness Visual Studio Code documentation
  17. Manage AI settings in enterprise environments Visual Studio Code documentation
  18. Introducing the Microsoft Copilot harness Microsoft

Frequently asked questions

Which GitHub Copilot models were deprecated on 16 November 2026?

Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. GitHub suggests Claude Opus 5.5 in place of both Opus models, and GPT-6.1 Sol or GPT-6 Luna in place of GPT-5.6 Terra. Claude Opus 4.8 remains available to individual Copilot subscribers on annual plans.

Do Copilot Business and Enterprise admins need to enable the replacement models?

Usually not. Under default model enablement the suggested alternatives are turned on automatically unless an administrator has turned off the global default or explicitly disabled the model. If the global default is off, enable the alternatives through their model policies in Copilot settings.

Which Copilot models are being deprecated in December 2026?

On 14 December: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3, with suggested replacements Fable 5.2, Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon, Grok 4.7 and one to be announced. MAI-Code-1.1-Flash goes on 18 December. Haiku 4.5 stays for individual annual-plan subscribers.

When is the VS Code Local harness removed?

In February 2027. VS Code 1.146 deprecated it, and the GitHub changelog of 16 November 2026 describes a three-month transition. The vscode.temporaryDefaultToLocalHarness enterprise-managed setting is removed at the same time, so keeping it configured will not keep the Local harness available.

Does switching to the Copilot harness move my code to the cloud?

No. GitHub says choosing the Copilot harness doesn't move your files to the cloud, and agent commands still run on your machine. It also does not turn on sandboxing; administrators and users configure sandboxing separately.

Will my prompt files work in the Copilot harness?

No. VS Code's migration guide says Copilot harness sessions do not load *.prompt.md files and use agent skills instead. The migration list in Chat: Open Customizations can convert workspace and user prompt files to skills; recheck any agent, model or tool settings the conversion drops.

What does strictPluginOnlyCustomization do?

It is an enterprise-managed setting that blocks standalone customizations from user and project sources, so they must come from eligible plugins or managed settings. It can apply to all five categories, skills, custom agents, hooks, MCP servers and executable Copilot extensions, or to selected ones. GitHub says to pair it with strictKnownMarketplaces to restrict which marketplaces count.

All tools →