Two specific models were named in GitHub's 31 July deprecation notice: Gemini 2.5 Pro and Gemini 3 Flash.
The replacement was already there. Gemini 3.6 Flash landed in Copilot on 21 July, ten days before the notice.
A generation lasts about a year
Gemini 2.5 Pro was, not long ago, a headline model. Gemini 3 Flash is newer than that and is also going.
This is the pattern to internalise: at current cadence, the model you standardise on today has roughly a year before it is deprecated, and less than that before something clearly better exists.
What to check if you used them
- Any prompt tuned specifically for Gemini 2.5 Pro's long-context behaviour — re-test it, do not assume it transfers
- Anything routing to Gemini 3 Flash for cost reasons — 3.6 Flash is the successor, but verify the cost maths rather than assuming parity
- Evaluation baselines recorded against either model, which are now unreproducible
The structural point
This is the same week OpenAI cut GPT-5.6 prices by up to eighty percent and Copilot added a fourth vendor. Capability rises, prices fall, and models disappear — all at once.
The defensible position is not picking the right model. It is building so that changing model is a configuration change rather than a project.