Run this 15-minute pinned-model audit before Copilot's October 2 deprecation breaks your workflow

A deprecation is a workflow audit

If you pinned a Copilot model, the audit is short and the date is fixed. GitHub is retiring Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 from GitHub Copilot on October 2, 2026. The deprecation applies across all Copilot experiences, including chat, inline editing, agent and ask workflows, and code completion. Users do not have to manually delete the models after the deprecation. That does not make the work optional. A pinned model can still point at a name that stops working, so the useful task is to find every pin, replace it, and check the new cost line.

The audit stays short with a known map of model pins. It tests whether your workflow can still run after the old names disappear, not whether the dead entries are tidy. A setting that looks harmless still needs a throwaway file test. A script with a model name needs the string replaced, then the smallest job that would have used it. The audit is not complete until the workflow can run without the retired name.

The checklist catches pinned models

Run the items in order. Each item is a place a model can be pinned, and each is quick to test. The order matters because a fix can fail in a place you do not check.

• Open the model picker and note any model from the retired set still selected.
• Check the org or enterprise model policy for retired model access and replacement enablement.
• Search CI scripts, dotfiles, and local config for retired model names.
• Review agent workflows for a pinned model in agent mode.
• Map each retired model to its named replacement.
• Re-price the replacement before the deadline.

Send the request now for any needed change. Prove a setting with a small test before you trust it. Keep a short note of what you changed: the old model, the new model, the file, the reason, and the owner. A failed replacement points you to the note. A working one becomes the next team's starting point.

Replacement mapping changes the cost line

Gemini 3.8 Flash replaces both Gemini Flash models, Kimi K3 replaces Kimi K2.7 Code, and Claude Opus 5 replaces Claude Opus 4.7. Kimi K3 has a 1M-token context window, four times the 256K limit of Kimi K2.7 Code. Gemini 3.8 Flash is priced at $0.75 per million input tokens and $3.75 per million output tokens, with that price doubling on January 1, 2027.

Before you trust the replacement, give it a task you already know. Use a small branch, a narrow file, and a prompt that used to work. Worse output gets a failure-mode note. Better output gets a note on what changed. The replacement earns its pin only after it survives your own baseline.

When you re-price the replacement, do not compare list prices alone. Count the tasks that run daily, the jobs that run nightly, and the agent sessions that run long. A model that looks cheap on a single prompt can become expensive in a loop. A model that looks expensive can be fine if it finishes in fewer turns. Flag the price doubling before it changes your plan.

Policy and scripts decide whether the fix sticks

Copilot Enterprise and Business administrators may need to enable the replacement models in org model policies before October 2, 2026. Policy is the blocker. A user can pick a replacement that is not enabled and still get blocked. Ask for the policy change in writing, with the model name and the date.

Scripts are the least forgiving part. A model name in a CI script can remain after a UI update, a repo rename, and a team handoff. A script failure after the deprecation may arrive in a queue, not in your face. The note you keep during the audit is the difference between a quick fix and a repeated incident.