GitHub Copilot is retiring Gemini 2.5 Pro and Gemini 3 Flash on July 31, 2026. The official announcement landed on the GitHub Changelog on July 2, and it affects every Copilot surface: Chat, agent mode, ask mode, inline edits, and code completions. GitHub’s recommended replacements are Gemini 3.1 Pro and Gemini 3.5 Flash.
For most individual developers this is a thirty-second fix. For teams — especially on Copilot Business or Enterprise plans, where model access is governed by policy — there are a few places this change can quietly break a workflow. This guide covers both, with a checklist you can run before the deadline.
Key takeaways
- Gemini 2.5 Pro and Gemini 3 Flash are removed from all GitHub Copilot experiences on July 31, 2026 — Chat, agent and ask modes, inline edits, and code completions.
- GitHub recommends Gemini 3.1 Pro (replacing Gemini 2.5 Pro) and Gemini 3.5 Flash (replacing Gemini 3 Flash).
- Individual users: switch via the model picker in Copilot Chat. No other action needed — retired models are removed automatically.
- Copilot Business and Enterprise admins: the replacement models may need to be enabled through model policies in Copilot settings before your developers can select them. This is the step most likely to catch teams out.
- Audit your team’s references: internal docs, onboarding guides, custom instructions, and prompt files that name the old models should be updated before the deadline.
- This only affects Copilot. Google’s Gemini API and AI Studio follow Google’s own model lifecycle, not GitHub’s.
Key facts at a glance
| Question | Answer |
|---|---|
| What was announced? | Retirement of two Gemini models from GitHub Copilot |
| Announced | July 2, 2026 (GitHub Changelog) |
| Effective date | July 31, 2026 |
| Retiring models | Gemini 2.5 Pro, Gemini 3 Flash |
| Recommended replacements | Gemini 3.1 Pro, Gemini 3.5 Flash |
| Scope | All Copilot experiences: Chat, agent mode, ask mode, inline edits, code completions |
| Admin action | Business/Enterprise: enable replacements via model policies if not already enabled |
| Individual action | Switch models in the Copilot Chat model picker |
Table of Contents
- What exactly is changing?
- Which models should you switch to?
- Do individual developers need to do anything?
- What Copilot Business and Enterprise admins need to do
- Where old model names hide in your workflow
- Why does GitHub retire Copilot models?
- What this means for engineering and QA teams
- Frequently asked questions
- Conclusion
What exactly is changing?
On July 31, 2026, GitHub removes two models from Copilot’s lineup:
- Gemini 2.5 Pro — Google’s previous-generation flagship, long a popular pick for complex reasoning tasks in Copilot Chat.
- Gemini 3 Flash — the low-latency option many developers used for quick questions and fast iterations.
The retirement applies across every Copilot experience: Copilot Chat on github.com and in IDEs, agent mode, ask mode, inline edits, and code completions. After the deadline, the models no longer appear in any model picker.
One reassuring detail from the changelog: “No action is required to remove the older models once they have been deprecated.” The removal is handled entirely on GitHub’s side — nothing in your editor or repository breaks by itself.
Which models should you switch to?
GitHub’s guidance maps each retiring model to its current-generation successor:
| Retiring model | Recommended replacement | Best for |
|---|---|---|
| Gemini 2.5 Pro | Gemini 3.1 Pro | Deep reasoning, multi-file refactors, complex agent tasks |
| Gemini 3 Flash | Gemini 3.5 Flash | Fast iterations, quick explanations, low-latency completions |
The mapping preserves the trade-off you already chose: if you picked the Pro tier for reasoning depth, move up to the newer Pro; if you picked Flash for speed, move to the newer Flash.
Two practical notes before you switch:
- Check the request multiplier. On paid Copilot plans, each model has a premium-request multiplier that affects how quickly you consume your monthly allowance. Verify the replacement’s multiplier on GitHub’s supported models reference so your usage math still holds.
- Re-run your judgment, not just your prompts. A newer model is not automatically better for your specific workflow. Give the replacement a week on real tasks before you standardize on it — the behaviors that matter (how it handles your codebase’s conventions, how verbose its diffs are) only show up in practice.
Do individual developers need to do anything?
If you use Copilot’s default model: no. The change is invisible to you.
If you deliberately selected Gemini 2.5 Pro or Gemini 3 Flash: switch before July 31 so the transition happens on your terms rather than mid-task. It takes seconds:
- VS Code / JetBrains / other IDEs: open the Copilot Chat panel and use the model picker (the dropdown showing the current model name near the chat input). Select Gemini 3.1 Pro or Gemini 3.5 Flash.
- github.com: open Copilot Chat and choose the replacement from the same model dropdown.
- Agent workflows: if you launch Copilot coding agent tasks with a specific model preference, update that preference the same way.
If you wait past the deadline, nothing catastrophic happens — the retired model vanishes from the picker, and in past Copilot retirements sessions simply fell back to the default model. But an unplanned model change mid-sprint is exactly the kind of variable you don’t want to discover while debugging.
What Copilot Business and Enterprise admins need to do
This is the part of the announcement that deserves the most attention, because it’s the only step that can silently block your developers.
Per the changelog: “Copilot Enterprise administrators may need to enable access to the alternative models through their model policies in Copilot settings.”
Concretely, before July 31:
- Open your organization or enterprise Copilot settings and review model policies.
- Confirm Gemini 3.1 Pro and Gemini 3.5 Flash are enabled. If a policy was configured back when Gemini 2.5 Pro was approved and never revisited, the newer models may still be switched off.
- Once enabled, the models appear in every developer’s Copilot Chat model picker automatically.
- Announce the change internally — one message in your engineering channel now prevents a wave of “why did my model disappear?” tickets in August.
GitHub also notes that Enterprise customers with questions can raise them with their account manager.
Where old model names hide in your workflow
The models will be removed cleanly, but references to them won’t be. Before the deadline, grep your team’s material for Gemini 2.5 Pro and Gemini 3 Flash:
- Onboarding docs and wikis — “set your Copilot model to Gemini 2.5 Pro” instructions will dead-end for new hires.
- Custom instructions and prompt files — files like
.github/copilot-instructions.mdor shared prompt templates that recommend a specific model. - Runbooks and demo scripts — recorded workflows, conference demos, internal training decks.
- Benchmark notes — if your team compared models for a specific task and documented “use 2.5 Pro for X,” re-validate that guidance against Gemini 3.1 Pro rather than deleting it.
None of these break at a technical level — they just rot into misinformation. A ten-minute search-and-update now keeps your documentation trustworthy, which is the whole point of having it.
Why does GitHub retire Copilot models?
Copilot is a multi-model product: GitHub curates a menu of models from several providers and retires older generations as newer ones prove themselves. Keeping every historical model available forever would mean maintaining, securing, and paying for inference on models that newer ones outperform — complexity that delivers no benefit to users.
This isn’t a one-off event, either. GitHub retired earlier Gemini and GPT models in March 2026 on the same pattern: changelog announcement, a few weeks’ notice, a recommended successor. If your team builds Copilot into its daily workflow, model retirements are now a recurring operational event — roughly like a dependency update, except the vendor picks the date.
What this means for engineering and QA teams
There’s a broader lesson here that goes beyond this one announcement, and it’s the same one we drew from Claude Fable 5’s sudden three-week suspension: the AI model in your workflow is a dependency with an availability lifecycle you don’t control.
Treat it like any other dependency:
- Keep an inventory. Know which models your team actually uses and where they’re named — the same way you’d track a library version. When a retirement lands, the audit is a lookup instead of an archaeology project.
- Don’t couple process to a model name. Write internal guidance in terms of tiers and tasks (“use the fast model for boilerplate, the reasoning model for refactors”) so a retirement changes one mapping table, not fifty documents.
- Verify after switching. Model swaps change output subtly — different assertion styles in generated tests, different diff verbosity, different defaults. Our review of a study on generative AI in software testing found the failure mode isn’t dramatic breakage; it’s plausible-looking output that’s quietly wrong. Whatever model writes your code, the review and test discipline around it is what actually protects your codebase.
Frequently asked questions
When will Gemini 2.5 Pro and Gemini 3 Flash stop working in GitHub Copilot?
On July 31, 2026. GitHub announced the retirement on July 2, 2026, and it applies to all Copilot experiences: Chat, agent and ask modes, inline edits, and code completions.
Which models replace Gemini 2.5 Pro and Gemini 3 Flash in Copilot?
GitHub recommends Gemini 3.1 Pro as the replacement for Gemini 2.5 Pro and Gemini 3.5 Flash as the replacement for Gemini 3 Flash — the current-generation successors in the same capability tiers.
What happens if I still have a retired model selected after July 31, 2026?
The retired models are removed from the model picker automatically; GitHub states no action is required to remove them. In past Copilot model retirements, affected sessions fell back to the default model. To stay in control of the switch, select your replacement before the deadline.
Do Copilot Free, Pro, or Pro+ users need to do anything before the deadline?
Only if you actively selected one of the retiring models. Open the Copilot Chat model picker and choose Gemini 3.1 Pro or Gemini 3.5 Flash. If you’ve always used the default model, this retirement doesn’t affect you.
Why can’t I see Gemini 3.1 Pro or Gemini 3.5 Flash in my Copilot model picker?
On Copilot Business and Enterprise plans, individual models must be enabled through model policies in Copilot settings. If the replacements aren’t visible, your administrator hasn’t enabled them yet — once enabled, they appear in the picker automatically.
Does this retirement affect the Gemini API or Google AI Studio?
No. GitHub’s retirement only changes which models Copilot offers. Google’s own platforms — the Gemini API, AI Studio, and Vertex AI — follow Google’s separate model lifecycle and deprecation schedule.
Conclusion
This is a routine, well-telegraphed model retirement — but “routine” doesn’t mean “ignorable.” The thirty-second version of your to-do list:
- Switch now if you use Gemini 2.5 Pro (→ Gemini 3.1 Pro) or Gemini 3 Flash (→ Gemini 3.5 Flash).
- Admins: confirm the replacements are enabled in your Copilot model policies.
- Update any docs, prompts, or runbooks that name the old models.
- Re-validate any model-specific guidance your team documented.
Model retirements are becoming a normal part of AI-assisted development, on the vendor’s schedule rather than yours. Teams that treat their AI models like tracked dependencies — inventoried, abstracted, verified after every change — absorb these announcements in minutes. Teams that don’t, rediscover their whole AI workflow in production during the first week of August.
Sources: GitHub Changelog — Upcoming deprecation of Gemini 2.5 Pro and Gemini 3 Flash · GitHub Docs — Supported AI models in Copilot
