GitHub Kills Gemini 2.5 Pro and Gemini 3 Flash in Copilot Tomorrow. If You Pinned a Model Months Ago and Forgot, Check Today.
Tomorrow, July 31, GitHub removes Gemini 2.5 Pro and Gemini 3 Flash from every Copilot surface: Copilot Chat, inline edits, ask mode, agent mode, and code completions. GitHub announced this back on July 2, gave nearly a month of notice, and recommended Gemini 3.1 Pro and Gemini 3.5 Flash as the replacements. That's a reasonable amount of runway for a team with someone whose job includes reading platform changelogs.
It's a much worse amount of runway for a solo operator, because there's no second person whose job that is. You're the developer, and you're also the platform admin who was supposed to see the July 2 notice, and if those two roles didn't sync up three weeks ago, tomorrow is when it shows up as a broken workflow instead of a calendar reminder.
What's actually ending
Gemini 2.5 Pro and Gemini 3 Flash are being deprecated across the full range of Copilot experiences (chat, inline code edits, ask mode, agent mode, and code completion) as of July 31, 2026. This isn't a quiet model swap where a newer version silently takes over the same slot; it's a hard cutoff on two named models across every surface that could reference them. GitHub's own recommendation is to move to Gemini 3.1 Pro or Gemini 3.5 Flash depending on which one you were using, and for Copilot Enterprise, administrators may need to explicitly enable access to the replacement models through their model policies before anyone on the team can use them.
Why this hits differently for a one-person operation
An enterprise team has a defined role (a platform admin, a DevEx lead, someone whose actual job function includes tracking model deprecations and updating org-wide policy) who owns this kind of change. A solo operator doesn't have that role as a separate person. You set the model in a moment of focus three weeks or three months ago, moved on to the actual work you were trying to do, and the changelog entry that announced this on July 2 is one of dozens of similar notices you've seen and mentally filed under "not today." Today is the day it becomes today.
Where model pins actually hide
The places worth checking, in order of how easy they are to forget about: any repo-level .github/copilot-instructions.md or custom-instructions file that names a specific model; VS Code or JetBrains extension settings where you or a past collaborator explicitly selected a model instead of leaving it on default; any CI pipeline or script that invokes Copilot CLI or agent mode with an explicit --model flag; and any onboarding note, internal doc, or README that tells a future collaborator (including future you) which model to pick when starting a new session.
The five-minute check
Open your Copilot settings and search for "2.5 Pro" or "3 Flash" across the places listed above. If you find either name hardcoded anywhere, swap it to Gemini 3.1 Pro or Gemini 3.5 Flash respectively before end of day. If you're on Copilot Enterprise or manage model policy for a small team, confirm the replacement models are actually enabled in your org's settings: GitHub's own guidance flags that this isn't automatic for enterprise accounts.
The habit worth building instead
This is the fourth or fifth model deprecation across GitHub, OpenAI, Anthropic, and Google in the past several months, and the cadence isn't slowing down: if anything, monthly model churn across every vendor is becoming the default rhythm of building on frontier AI in 2026. Reading every changelog isn't a sustainable defense against that; you will miss one eventually; changelogs are exactly the kind of thing that competes with actual shipping and loses. The more durable fix is never hardcoding a specific model version anywhere you won't personally look again for months. Default configs, unpinned model references, or a single config variable you update in one place: anything that turns "GitHub deprecated a model" into a five-second fix instead of a scavenger hunt through every file that might reference it.
What I'd actually do
Run the five-minute check today, then spend fifteen more minutes finding every hardcoded model reference across your setup and replacing it with a default or a single named constant you control in one place. That second step is the one that actually stops this from happening again next quarter, when it's a different model on a different vendor's deprecation list.
Author
Lukas
@lukcombinator