GitHub Copilot deprecated six model choices on October 19

This was a Copilot deprecation, not a shutdown of the models everywhere

On October 19, GitHub confirmed that six model selections were deprecated across most GitHub Copilot experiences. The list is Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5. GitHub names Copilot Chat, inline edits, ask and agent modes, and code completions among the affected surfaces. Removing the retired choices requires no action, though Enterprise administrators may need to enable a replacement through policy.

The word “deprecated” here is service-specific. GitHub is saying these model choices are no longer supported in most Copilot experiences; it is not announcing that Google, OpenAI, or xAI has shut down the underlying model everywhere. A developer who uses one of them directly through another provider should not infer anything about that separate API from the Copilot notice.

The final notice gives different replacement suggestions

GitHub’s October 19 table suggests Gemini 3.8 Flash for Gemini 3.7 Flash, GPT-6.1 Sol for GPT-5.5 and GPT-5.4, GPT-6 Luna for GPT-5.4 mini and GPT-5 mini, and Grok 4.7 for Grok 4.5. That is useful migration guidance, not a guarantee that every user can select the replacement immediately or that it behaves identically on the same prompt. GitHub’s current supported-model table marks all four suggested alternatives as generally available, a release status that remains separate from account access.

The September 18 advance notice listed different GPT and Grok alternatives at the time. That is a reminder to use the final retirement post and current model list rather than copy a replacement from an earlier announcement into a workflow. Model lineups can change while a deprecation is pending.

A supported model may still be missing from a particular selector

GitHub’s supported-model documentation says model access depends on plan and client. Copilot Free and Student users get models through Auto selection only. The page separates its general model list, Auto list, client table, and plan table, so an entry in the main list does not promise that every user can select it in every interface. Check the table for the plan and mode developers actually use. In Visual Studio Code, users can add models beyond the default list when their Copilot subscription offers them.

Client versions can add another gate. GitHub’s current minimum-IDE table lists Visual Studio 17.14.19 for GPT-6.1 Sol and TBD for Visual Studio Code. The documentation explains that a missing minimum means the release guidance it reviewed did not specify one; it does not certify every older release. The table is tentative, and GitHub recommends keeping the IDE and Copilot extension or plugin current. Record the client version when deciding whether a replacement is ready for a rollout.

Business and Enterprise policies can hide a model that the general documentation lists. GitHub’s October 19 model-policy API announcement says GET and PATCH endpoints are available for enterprise, enterprise-team, agent-app, and organization policies. Enterprise policy can delegate control to organizations in organization mode, or to teams and apps in enterprise teams mode. That setting tells administrators where to inspect access and who can enable the replacement. GitHub gives each scope a separate endpoint; use the one matching the delegated owner and read its current settings before applying a change.

GitHub’s model-policy documentation says the default-availability policy for Copilot Business and Enterprise applies to generally available models that have not been configured individually. A newly released GA model inherits that default until an administrator configures it explicitly. Pre-GA and open-weight models, models outside GitHub’s data-retention agreement, and models that do not meet a data-residency or FedRAMP restriction are excluded and disabled by default. Before changing policy, check whether the replacement is covered by the default or needs an explicit rule.

Auto-selection and pinned models need different checks

A pinned choice can disappear from a selector or command option after retirement. Auto may keep working while choosing from its own supported subset, which can change as GitHub updates the list. GitHub’s supported-model documentation names GPT-5.3-Codex as the long-term-support fallback when no other Auto models are available. A team that relies on Auto should therefore check both the model policy and the fallback it could receive.

GitHub also separates generally available choices from evaluation models. Evaluation models may appear under codenames, can be added or removed without notice, and may have different availability or rate limits. GitHub says its tests found that they may perform worse on security-related prompts. For individual plans, a policy controls whether Auto can use them; users can disable them under Settings, AI controls, and Evaluation models in Copilot auto model selection. If Auto results shift without a listed replacement, check that setting before attributing the change to a model migration.

Make one migration inventory before editing anything: search scripts, prompts, repository instructions, onboarding notes, and evaluation records for the six retired names, then mark each reference as fixed or Auto. Add the Copilot client and plan beside it. For a fixed reference, choose from GitHub’s final replacement table only after checking client support, version, and policy. For Auto, record that the routed model can change and confirm the allowed subset in the actual client. This gives the team one list to review instead of repeating the same search in each project document.

Test the new choice on the task that mattered

Before changing a team default, reuse one bug fix and one code review that expose why the old model was chosen. Run them in the target Copilot client with the same repository state, prompt, tools, and instructions. Compare task success, the corrections a reviewer has to make, and whether the workflow follows its existing permission and approval steps. A result from Chat does not establish the behavior of code completion or an agent, so include the surface the team depends on.

GitHub’s model documentation describes extended capabilities such as a one-million-token context window and configurable reasoning for supported models. The larger context is available only in Visual Studio Code and Copilot CLI; configurable reasoning is available in Visual Studio Code, Copilot CLI, and Copilot cloud agent. Higher settings consume more AI credits, and GitHub recommends regular settings by default. Keep those settings constant when comparing two models; test extended context or reasoning separately if the task needs it, then account for the added credits.

Keep a short result record with the model, date, client and version, plan or policy, repository revision, task prompt, and outcome. Include reviewer corrections and any change to instructions or tool permissions.