If you pinned Gemini 2.5 Pro or Gemini 3 Flash anywhere in GitHub Copilot, both are gone as of July 31, 2026. GitHub deprecated them across every Copilot experience — chat, ask and agent modes, inline edits, and code completions (GitHub Changelog). The removal is automatic and needs nothing from you. Re-pointing everything that named those models is the part that's on you, and it takes about ten minutes if you know where to look.

Here's the answer up front: move 2.5 Pro → Gemini 3.1 Pro, and 3 Flash → Gemini 3.5 Flash (GitHub's named successors), or jump 3 Flash straight to Gemini 3.6 Flash, the cheaper workhorse tier added to Copilot on July 21. The rest of this is the checklist so nothing silently falls back to a model you didn't choose.

1. Find every place the old model is named#

A model id leaks into more configs than people remember. Check, in order:

Anything still naming a deprecated model won't error into your face; it drops to a default. That silent fallback is the whole risk.

2. Pick the replacement — and use the moment#

The direct swaps are in the table above. But a forced migration is the cheapest possible time to re-evaluate, because you're already in the config. A pin you set months ago and never measured is exactly the thing to A/B now. Re-run a small eval on cost per completed task, not tokens-per-second or the benchmark you half-remember — a cheaper model that needs one more retry costs more. If your Copilot exposes non-Gemini options, this is a free chance to put them side by side before you re-pin.

3. The admin step that actually unblocks people#

On Copilot Enterprise and Business, models are gated by an org model policy. A replacement only appears in the Copilot Chat model selector in VS Code and on github.com after an admin enables the policy for it. So the sequence that strands teams is: the old model retires, no replacement policy is enabled, and everyone quietly loses Gemini access with no obvious cause.

If you're the admin: open Copilot settings, enable the policy for Gemini 3.1 Pro and/or the Flash tier you want, and confirm it renders in the selector. If you're not: this is the one thing to ask your admin for by name, because you cannot enable it yourself.

4. Verify, then move on#

Re-open the model picker and confirm your chosen replacement is selectable, run one real task through it (a chat, an agent run, a completion in the file type you actually work in), and check that nothing in CI still passes a dead model id. That's the loop: find the pins, pick on cost-per-task, flip the admin policy, verify. Ten minutes, and you're off a deprecated model instead of discovering it mid-sprint.

The broader signal: your coding agent's default is now a moving part. Between this retirement, the week's model launches and price cuts, and open-weight options landing in Copilot, "set it once" isn't a strategy anymore. Keep a two-line eval you can re-run, and treat every deprecation notice as a scheduled benchmark.