Four GitHub Copilot models retired: replacing Opus 4.7 and preparing for October 19

Two dates and ten names: group by deadline first

GitHub's October 2, 2026 confirmation says Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 are retired. The first two map to Gemini 3.8 Flash, Kimi to Kimi K3, and Opus to Claude Opus 5.5. This newer confirmation updates the Opus 5 mapping in the September 3 advance notice; use the latest recommendation when migrating.

On September 18, GitHub separately announced six October 19 retirements: Gemini 3.7 Flash → Gemini 3.8 Flash; GPT-5.5 and GPT-5.4 → GPT-5.6 Sol; GPT-5.4 mini and GPT-5 mini → GPT-5.6 Luna; Grok 4.5 → Grok 4.6. Arrows show recommended replacements, not measured equivalence, automatic configuration rewrites or simultaneous permissions for everyone. Track each round separately so the later date is not applied to the earlier models.

Retirement is confirmed; now check your own configuration

The October 2 announcement changes the event from planned deprecation to confirmed retirement. As of October 3 Beijing time, the official overall supported-model table reviewed here no longer lists the four models, and retirement history records 2026-10-02. This article therefore removes its earlier observation that the support table still listed them. Official status and an organization's migration are separate: no organization login, selector check or call test was performed here, so this does not prove any organization completed its switch.

Record plan, organization, Copilot surface, IDE/extension version and whether model names are pinned. Administrators should check each replacement's model policy in Copilot model settings, then users should inspect selectors in supported clients. GitHub says retired models do not need manual removal but still requires workflow and integration updates; this does not mean every pinned configuration repairs itself.

If an old name remains visible, treat it as client state needing investigation, not an extension or guarantee of continued service. If a replacement is absent, check plan, client support and organization policy. If a pinned name fails, identify the effective configuration location. Preserve surface, version, time and original error, then test availability with a small task without external writes. For simultaneous failures across vendors, also check GitHub Status for an outage.

Inspect every Copilot experience, beyond the model selector

Both GitHub announcements concern Copilot experiences including Copilot Chat, inline edits, ask/agent mode and code completions. Also inspect GitHub.com, IDEs, Copilot CLI, cloud agent, Copilot app, repository instructions, scripts and internal templates for pinned names. Actual client model availability still varies by plan, version and policy. Do not extend GitHub's GPT-5.5 notice into a same-day ChatGPT, Codex or OpenAI API change.

Unattended jobs are easy to miss: interactive users change models while schedules or shared settings retain old references until the deadline. Inventory old model, date, recommended replacement, surface, configuration location, owner, last successful run and verification state. This is a preventive recommendation, not a failure observed by this site.

Business and Enterprise: check model policies before user migration

For the October 19 list, GitHub says recommended replacements automatically enable for Business and Enterprise using default model enablement, except when administrators disable the global default or a specific model. Check Gemini 3.8 Flash, GPT-5.6 Sol, GPT-5.6 Luna and Grok 4.6 individually in Copilot model policies. Separately check the October 2 replacements Gemini 3.8 Flash, Kimi K3 and Claude Opus 5.5.

No organization administrator login or account test was performed here. Public support does not guarantee a particular plan, client and organization policy has access. Administrators should inspect their own model policies, then authorized users should check selectors in their actual clients. Personal-account visibility cannot establish organization-wide access.

Check Kimi K3 separately. Current official default-enablement scope excludes open-weight models such as Kimi K3, which default off regardless of global Default availability. A global enabled setting does not substitute for Kimi K3's specific approval; verify explicit activation under the organization's actual policy.

Replacement is not equivalence: compare tasks and costs first

Official recommendations indicate candidates, not equivalent output, latency or tool behavior for your team. Fix representative tasks and compare code diffs, tool calls, human rework, latency and errors in a controlled scope. In particular, do not assume GPT-5.4 mini and GPT-5 mini become identical in cost or capability when replaced with Luna. This site ran no benchmarks, so this article ranks no model performance.

GitHub model and billing tables change. Check the actual plan, usage units and current rates at migration time. Preserve old/new prices, organization AI-credit snapshots and actual selected-model usage before budgeting. The October 2 Kimi replacement may also change costs. Publication-time prices are not a permanent contract, and switching models does not imply unchanged charges.

Migrate and verify each deadline separately

First search the four October 2 and six October 19 old names separately, recording surfaces and configuration precedence. Second have administrators check recommended replacement policies. Third change one model at a time in testing, without simultaneously changing prompts, permissions and networks. Fourth use identical samples for answers, completion, edits and agent tool chains. Fifth compare AI credits, latency, failure rates and human rework. Sixth switch by team or repository, completing an unattended run before each respective deadline.

For agents with external writes, cover reads, one write with an idempotency key, tool timeouts, permission denial, cancellation and reconciliation after unknown outcomes. Stop expansion immediately for repeated submissions, missed approvals, repository-rule violations or excessive costs. Do not experiment with more production tasks.

Failure, stopping and fallback: old models are not permanent rollback targets

The four models already retired October 2 should no longer be rollback targets. Do not treat temporary availability of October 19 models as a long-term guarantee either. Preserve prompt and tool-configuration snapshots and prepare another supported, policy-compliant model verified on representative tasks. Roll back to a reviewable working solution, not a withdrawn product.

Pause the path if replacements remain absent, policies repeatedly override local settings, critical quality regresses, AI credits exceed budget or agents duplicate side effects. Retain client, configuration source, model, time, original error and run ID. Do not automatically cycle through multiple models and lose attributable evidence.

Network boundaries and follow-up: retirement is not a connection failure

A model disappearing, organizational disabling, an invalid pinned name or unavailable replacement belongs to Copilot's product and permission layers. Changing IPs or proxies cannot restore it. Investigate networks only for independent DNS failures, TLS handshake errors, proxy 407, connection timeouts or unreachable services. PuppyIP's proxy connection failure checklist can help with those independent issues.

Before each deadline, recheck GitHub Changelog, supported models, client support and organizational policies. Update this same URL if GitHub changes dates, mappings, prices or scope. Without new official facts, one account's display cannot establish worldwide progress. This is a migration checklist, without organization-account login or production switching.