github copilot drops gemini 2.5 pro and gemini 3 flash

GitHub deprecated Gemini 2.5 Pro and Gemini 3 Flash across all Copilot experiences on July 31, 2026. The deprecation covers Copilot Chat, inline edits, ask mode, agent mode, and code completions — every surface where either model appeared. Both models are no longer available for selection.

what replaces what

The replacement for Gemini 2.5 Pro is Gemini 3.1 Pro, currently in public preview. Gemini 3 Flash is replaced by Gemini 3.6 Flash, which became generally available in Copilot on July 21, 2026 — ten days before the older model was removed. GitHub gave advance notice on July 2, giving teams approximately four weeks to update workflows and test replacements.

Gemini 3.6 Flash carries a spec improvement over Gemini 3 Flash: a 1M token context window, a 64K output cap, and pricing 17% lower than Gemini 3.5 Flash at $1.50/$7.50 per million tokens (input/output). For code completion and fast iteration workflows, the upgrade should be transparent — the model is faster and cheaper at the same context length.

Gemini 3.1 Pro as a replacement for 2.5 Pro is more significant. Gemini 3.1 Pro is still in public preview, which means it may not be enabled by default in all Copilot plans. Enterprise admins who relied on Gemini 2.5 Pro and have not yet enabled Gemini 3.1 Pro in their model policy may find the model unavailable until they explicitly turn it on.

enterprise model policy: new team-level controls

The deprecation landed on the same day as a separate but related announcement: enterprise teams model policy targeting entered public preview on July 31, 2026, with rollout beginning August 3.

Previously, Copilot Enterprise admins could set model access at the organization level. The new team-level targeting adds a layer: admins set a baseline for the entire enterprise, then grant specific models to individual enterprise teams. The model states are Enabled (available to all members), Disabled (unavailable to anyone), and Optional (available for assignment to specific teams).

The practical use case is role-based access: frontier models like Gemini 3.1 Pro or Opus 5 enabled for senior engineers or early experimentation teams, while standard teams get a more constrained set. It also solves a specific problem the Gemini 2.5 Pro deprecation created: if Gemini 3.1 Pro is in preview and you want to roll it out cautiously, you can now do that at the team level without either blocking everyone or enabling it for the whole enterprise at once.

what teams need to do

If your Copilot workflows referenced Gemini 2.5 Pro or Gemini 3 Flash explicitly — in API calls, in model policy rules, or in agent configurations — those references are broken as of July 31. The models will not route to a fallback automatically; they are unavailable.

For Gemini 3 Flash → Gemini 3.6 Flash: this is largely a drop-in upgrade. Update the model string and verify outputs, particularly for long-context tasks where the 1M context window becomes relevant.

For Gemini 2.5 Pro → Gemini 3.1 Pro: verify the model is enabled in your enterprise model policy before updating workflow references. If your enterprise model policy is still in its pre-team-targeting configuration, check the Copilot settings under the new enterprise teams mode toggle to see whether Gemini 3.1 Pro appears as an enabled option.

the model churn pattern

Copilot's model roster has been changing faster than most enterprise tooling budgets for. Since June 2026, Kimi K2.7 was added and deprecated, Gemini 3.6 Flash was introduced, Gemini 2.5 Pro and 3 Flash were removed, and Gemini 3.1 Pro entered preview — roughly one roster change every two weeks. For individual developers using Auto mode (where Copilot picks the model), this is invisible. For teams that have pinned specific models in enterprise policies, agent configs, or evaluation benchmarks, it requires active maintenance.

The new team-level policy targeting is partly a response to this problem: it lets enterprise teams experiment with newer models without forcing admins to make binary all-or-nothing decisions at each roster change. Whether four weeks' advance notice is enough for large organizations with complex Copilot integrations is a separate question — and one worth raising with your GitHub account team if your workflows depend on model-specific outputs.