update · TowCue Editorial Team
GitHub Copilot is retiring four models on October 2: how to protect your coding workflows
GitHub Copilot will retire Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 on October 2, 2026. Here is how teams should audit model dependencies before the cutoff.
Quick answer
GitHub announced on September 3, 2026 that four models will be deprecated across GitHub Copilot experiences on October 2, 2026:
| Model being retired | GitHub's suggested alternative |
|---|---|
| Gemini 3.5 Flash | Gemini 3.8 Flash |
| Gemini 3.6 Flash | Gemini 3.8 Flash |
| Kimi K2.7 Code | Kimi K3 |
| Claude Opus 4.7 | Claude Opus 5 |
GitHub says the change applies across Copilot Chat, inline edits, ask and agent modes, and code completions. Business and Enterprise administrators may also need to enable the replacement models in Copilot model policies before users can select them.
The practical TowCue takeaway is simple: a model name should not be a hidden hard dependency in a coding workflow. Teams using fixed models in Copilot CLI, shared instructions, internal docs, agent templates or evaluation scripts should audit those references before October 2 rather than waiting for users to hit an unavailable-model error.
For context on one of the main replacements, see TowCue's Gemini 3.8 Flash workflow analysis. For broader rollout discipline, use the AI agent best-practices guide and the AI tool selection guide.
What exactly is changing?
GitHub's September 3 changelog is a scheduled deprecation notice, not an immediate shutdown. The four listed models remain available until the October 2 cutoff, subject to the user's plan, client and organization policies.
The replacements are not one-to-one guarantees of identical behavior. A newer model can differ in reasoning style, latency, token use, tool calling and output format even when it is recommended as the successor.
That means a migration should be treated as a workflow regression test, not a model-picker housekeeping task.
Where can a fixed-model dependency hide?
The obvious place is the Copilot model picker, but team dependencies can live elsewhere:
- Copilot CLI commands that specify a model explicitly
- shell aliases or environment variables such as a fixed Copilot model
- onboarding docs telling developers to select one specific model
- prompt libraries tuned to the behavior of one model
- internal agent templates and evaluation suites
- screenshots or runbooks that assume an old model is available
- organization or enterprise policies that have not enabled the replacement
GitHub's documentation explicitly notes that model availability changes over time and depends on the Copilot plan and client. That is why a workflow that works for one developer can fail for another even before a deprecation date if policy or plan access differs.
Why should Business and Enterprise admins check policies now?
GitHub says Copilot Business and Enterprise administrators may need to enable replacement models through model policies.
This matters because GitHub's Default availability for released models policy determines whether newly released GA models inherit an enabled or disabled state when they have not been explicitly configured. Some classes of models are excluded from automatic enablement, including pre-GA models, open-weight models and models outside certain data-retention or compliance policies.
So the question is not only "Is the replacement model available in Copilot?" It is also:
"Is it available to this organization, under this policy, in this client?"
An admin should verify the answer before asking a team to migrate.
Should you switch to the suggested replacement immediately?
Usually, test first.
For Gemini 3.5 Flash and 3.6 Flash, GitHub recommends Gemini 3.8 Flash. GitHub's current pricing reference lists Gemini 3.6, 3.7 and 3.8 Flash in the same current price band, while 3.8 is the newest versatile Gemini option. But equal list pricing does not guarantee equal cost per completed task because model behavior and token consumption can differ.
For Claude Opus 4.7, GitHub recommends Claude Opus 5. For Kimi K2.7 Code, it recommends Kimi K3. Those migrations should also be tested against the exact coding jobs the team performs rather than accepted only because the successor is newer.
A good replacement passes your own acceptance tests on correctness, tool use, latency and review burden.
What about Copilot Auto model selection?
GitHub's Auto model selection can reduce some operational dependence on a single manually selected model. GitHub says Auto chooses from supported models based on availability, system health, performance, subscription and policy constraints.
For paid plans, GitHub also currently documents a 10% discount on model costs when Auto is used in supported Copilot surfaces.
But Auto is not a substitute for testing. If your workflow depends on deterministic behavior from one model, changing to Auto can introduce a different kind of variability. Use Auto when resilience and availability matter more than strict model consistency; use a fixed model when you have a tested reason to do so and a documented fallback.
A 30-minute migration audit
Before October 2, a small team can do a useful audit without redesigning its stack:
- Search repositories, scripts and internal docs for the four retiring model names.
- Check Copilot Business or Enterprise model policies for the suggested replacements.
- Pick 10–20 representative coding tasks that currently use the old model.
- Run the same tasks on the suggested replacement.
- Record accepted result rate, retries, latency and human correction time.
- Update fixed references only after the replacement passes the task set.
- Document a fallback model or use Auto where strict model pinning is unnecessary.
- Recheck the workflow before October 2, 2026.
This is especially important for agentic coding tasks. A model change can affect not just the final answer but also how the agent plans, calls tools, edits files and recovers from failures.
Who needs to act first?
Teams using Gemini 3.5 Flash or 3.6 Flash in Copilot
Act now if those models are explicitly selected in shared workflows. Gemini 3.8 Flash is the suggested successor, and TowCue has already covered the 3.8 Flash coding and agent tradeoffs.
Teams using Claude Opus 4.7 for difficult coding work
Regression-test Opus 5 on your hardest tasks. Do not assume identical instruction following or tool behavior just because GitHub recommends it.
Copilot Business and Enterprise administrators
Check policy availability before developers start migrating. A technically supported model can still be unavailable to users because of organization policy, plan or compliance settings.
Individual users who always use Auto
The immediate risk is lower because Auto selects from currently supported models. Still, remember that the model behind a response can change; review production code rather than treating Auto as a guarantee of behavioral consistency.
What should you not do?
Do not wait until October 2 and then replace every model string blindly.
Do not assume a successor will have identical cost, latency or prompt behavior.
Do not enable a new model organization-wide without checking data, compliance and policy requirements.
And do not build a long-lived agent workflow around a model name without a fallback plan. GitHub's own supported-model documentation explicitly warns that model availability is subject to change.
TowCue take
This deprecation is small as a product announcement but useful as a workflow warning.
AI coding stacks are becoming multi-model systems with short model lifecycles. The operational problem is no longer just choosing the best model today. It is designing a workflow that survives when today's model is renamed, replaced, repriced or removed.
TowCue would treat the October 2 cutoff as a reason to establish three habits:
- Keep model choice configurable instead of buried in code or documentation.
- Maintain a small regression set of real coding tasks for every important workflow.
- Define a fallback — either a second tested model or Auto selection where variability is acceptable.
That turns model retirement from an emergency into a routine dependency update.
Research sources
- GitHub Changelog — Upcoming deprecation of selected GitHub Copilot models (September 3, 2026)
- GitHub Docs — Supported AI models in GitHub Copilot (checked September 5, 2026)
- GitHub Docs — About default availability of Copilot models (checked September 5, 2026)
- GitHub Docs — About Copilot auto model selection (checked September 5, 2026)
- GitHub Docs — Models and pricing for GitHub Copilot (checked September 5, 2026)