GitHub Copilot retires Gemini 2.5 Pro and Gemini 3 Flash: the suggested replacement is still in preview

GitHub Copilot Retired a Production AI Model and Pointed Developers to a Preview Replacement

On July 31, GitHub removed two Google models from every surface in Copilot: chat, agent mode, inline edits, ask mode, and code completions 1. The models are Gemini 2.5 Pro and Gemini 3 Flash. GitHub's recommended replacement for Gemini 2.5 Pro is Gemini 3.1 Pro, which GitHub itself labels as a preview release 1.

A generally available model swapped for one still in preview. That gap is the story.

Twenty-Nine Days to Move

GitHub announced the deprecation on July 2, 2026 2. Twenty-nine days later, the models were gone. Copilot Enterprise administrators may need to enable access to the replacement models through their policy settings before developers can select them in the Copilot Chat model selector 1. GitHub noted that no action is required to remove the deprecated models from existing setups, and that Enterprise customers with questions should contact their account manager 1.

The Recommended Replacement Changed Mid-Notice

Between the July 2 announcement and the July 31 deprecation, GitHub updated both recommended replacements. The original announcement listed Gemini 3.5 Flash as the successor to Gemini 3 Flash, and Gemini 3.1 Pro for Gemini 2.5 Pro with no preview designation 2. By July 31, the Flash replacement had moved to Gemini 3.6 Flash, and the Pro replacement had picked up a preview label 1.

The reason is visible in the timeline. Gemini 3.6 Flash shipped on July 21, 2026 3, nineteen days after the deprecation was announced. GitHub updated its recommendation to the newer model, and the deprecation followed ten days later. A team that began evaluating Gemini 3.5 Flash on July 3 would have been preparing for a model GitHub no longer recommends.

Google's API Runs on a Longer Clock

Google's deprecation page lists the API shutdown date for gemini-2.5-pro as October 16, 2026 3. GitHub pulled the same model from Copilot on July 31, more than two months earlier. Teams calling Gemini 2.5 Pro through Google's API directly still have access. Teams using it through Copilot do not.

Shelf Lives Are Compressing

Google released Gemini 2.5 Pro as a generally available model on June 17, 2025 3. It lasted thirteen months in Copilot before deprecation. Gemini 3 Flash shipped on December 17, 2025 3 and was deprecated seven months later.

The replacement for Gemini 2.5 Pro, Gemini 3.1 Pro Preview, arrived in February 2026 3. Google's catalog lists it in the preview tier, separate from stable (GA) models, with no shutdown date announced yet 3. The replacement for Gemini 3 Flash arrived ten days before the deprecation. Teams are being asked to migrate from a model with a known support window to replacements that are either unlabeled for shutdown or brand new.

What This Costs Engineering Teams

The burden is not the models themselves. GitHub is redirecting which underlying model handles requests, not removing Copilot functionality. The cost is the re-validation work that redirect forces on engineering teams. Different models produce different output for the same prompts. Code completions, agent workflows, and inline suggestions calibrated to one model's behavior may produce different results with its replacement. For teams moving off Gemini 2.5 Pro, the replacement carries a preview label, a designation that signals the model's behavior may still change.

Google has shipped at least five model variants in the Gemini 3 family since November 2025: Gemini 3 Pro Preview, Gemini 3 Flash Preview, Gemini 3.1 Pro Preview, Gemini 3.5 Flash, and Gemini 3.6 Flash 3. Each release narrows the useful life of the one before it. Twenty-nine days is the full window GitHub gave teams to move off two models and onto replacements that were still shifting.

For teams building workflows around Copilot's model selection, the signal is structural. The model available today may be deprecated within weeks, and its replacement may still carry a preview label. Architecture that depends on a specific model being available is working against a release cadence that moves faster than quarterly planning.