Published November 15, 2026. A model-retirement deadline is a workflow-maintenance task, not just a reason to choose the newest name in a dropdown. This guide separates GitHub's announcement from an original readiness checklist. Nerova has not benchmarked these replacements.
Separate November 16 retirement
A separate Copilot retirement is scheduled for November 16, 2026. GitHub lists Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra; its suggested replacements are Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna. Opus 4.8 remains available to individual Copilot subscribers on annual plans. That November exception is separate from the Haiku 4.5 annual-plan exception in the December 14 wave below. Check the policy and model availability for your actual Copilot client and plan; the announcement does not specify an exact cutoff hour.
Which Copilot models retire on December 14?
GitHub's November 13 announcement schedules these changes for December 14, 2026 across Copilot experiences, including chat, inline edits, agent/ask modes and completions.
| Retiring model | Suggested replacement |
|---|---|
| Claude Fable 5 | Claude Fable 5.2 |
| Claude Haiku 4.5 | Claude Haiku 5.5 |
| Gemini 3.8 Flash | Gemini 4 Hydrogen |
| Gemini 4 Argon | Gemini 4 Carbon |
| Grok 4.6 | Grok 4.7 |
| Kimi K3 | Not yet announced |
Haiku 4.5 remains available to individual Copilot subscribers on annual plans, not all annual organizational contracts. GitHub handles removal; customers must update affected workflows. This is a Copilot notice, not a provider-wide API retirement. GitHub's announcement.
Enabled is not the same as migrated
The announcement describes automatic enablement of alternatives for Business and Enterprise under default policies, unless globally or individually disabled. But current supported-models documentation specifically excludes Fable 5.2 from default enablement and lists additional policy exceptions. Check actual settings rather than assuming every replacement is enabled.
GitHub's documentation also makes availability dependent on plan and client. Some combinations require minimum IDE or extension versions; a missing minimum does not establish compatibility with all older versions. Larger context and higher reasoning can increase credit consumption. Check supported models and policy exceptions.
Our recommendation: treat access, configured selection and accepted results as three separate checks. Seeing a model in a selector establishes neither that a pinned workflow uses it nor that its output satisfies your tests. Record all three states rather than closing a migration ticket after changing one setting.
Inventory actual work, not model popularity
Start with the repositories and workflows that matter to your team. For each, record its owner, client and version, selected model or automatic-selection mode, relevant policy and a repeatable acceptance check. Inspect the environment people actually use as well as documented configuration. Do not collect private prompts or code merely to populate an inventory.
| Field | Decision it supports |
|---|---|
| Owner and repository | Who can approve the change and evaluate output? |
| Client, version and selection | Where must availability and configuration be checked? |
| Task and acceptance criteria | What counts as useful work rather than plausible text? |
| Permissions and data constraints | Is the candidate permitted for this repository? |
| Continuity procedure | What happens if the selected route becomes unavailable? |
A small inventory covering consequential tasks is more useful than a comprehensive list nobody validates. Mark uncertain settings as unknown and assign a check. Do not turn an assumption into a green status.
Evaluate accepted changes, not first-answer speed
Use a disposable branch or isolated copy with approved, non-sensitive fixtures. Disable real deployments, external writes and payment actions. Select representative work: a narrow bug fix, a multi-file change, a test addition and a case with intentionally incomplete requirements. Hold the repository revision, task and acceptance criteria constant while comparing candidates.
Record whether each patch passes relevant tests, introduces unrelated changes, requires correction and respects the permitted scope. Include repeated attempts and human review time. A fast first answer that needs substantial repair can be less useful than a slower accepted result.
Define failure before running the comparison. Examples include a missing required test, an unrequested public-interface change, fabricated test results or an attempted side effect outside the allowed environment. These are proposed criteria, not claims about any particular model.
A cost-per-accepted-change worksheet
Divide total observed usage cost by accepted changes, reporting the acceptance definition and unsuccessful attempts alongside the ratio. Do not divide by generated answers and call that the cost of completed work. Keep human review effort separate unless you have an agreed valuation method.
For illustration only, a batch costing $18 that produces 12 accepted changes has a usage cost of $1.50 per accepted change. A $12 batch producing six accepted changes costs $2 per accepted change. These are hypothetical numbers, not GitHub prices or Nerova measurements. If no change is accepted, report no accepted outcome and the cost incurred, not a misleading zero.
A small sample can expose obvious breakages but cannot establish a universal winner. Repeat on representative tasks. Avoid changing the model, prompt and tool configuration together and then attributing the entire result to one variable.
A readiness gate before the deadline
- Confirm scope: identify affected selections and an accountable owner.
- Check access: inspect policy and the actual client without weakening security or data requirements.
- Run a bounded evaluation: retain the task revision, settings and acceptance result.
- Prepare continuity: document a separately validated supported option or manual workflow. A retiring model is not a dependable future rollback target.
- Roll out narrowly: validate a representative group before expanding and define when to stop on regressions.
- Recheck guidance: resolve outstanding replacement and policy questions before marking the migration complete.
When an alternative is unspecified, keep an owner on that dependency and assess only options your policies permit. Do not silently substitute another provider, expand permissions or promise untested equivalence.
Keep the workflow useful when models change
Model maintenance belongs in the operating cost of an AI workflow. Assign an owner to dependency changes, retain a small acceptance suite and keep a manual continuation path for important work. The objective is useful, maintainable execution rather than permanent attachment to a model name.
If you are still deciding what to automate, begin with a bounded process and measurable outcome. Nerova's operational-improvement project is a starting point for scoping that work, not an automatic Copilot migration service or a claim of an existing GitHub integration.