GitHub ships Gemini 3.5 Flash for Copilot
- GitHub said on May 19 it began rolling out Google’s Gemini 3.5 Flash in Copilot and put a new Copilot desktop app into technical preview. - GitHub said early testing showed “near-Pro coding quality” at Flash-tier speed and cost, with Gemini 3.5 Flash launching on a 14x premium request multiplier. - Starting June 1, 2026, GitHub plans to shift Copilot from request-based billing to usage-based billing, according to its pricing documentation.
GitHub said on May 19 that Google’s Gemini 3.5 Flash is now rolling out in GitHub Copilot, adding another model option to a product the company has been expanding across coding, review and agent workflows. In a changelog post, GitHub said the model delivered “near-Pro coding quality at Flash-tier speed and cost” in early testing and described it as suited to fast, iterative agentic coding workflows. GitHub also said a new Copilot desktop app entered technical preview on May 14, giving users a GitHub-native client to start work from repository context and move changes through pull request review. ### What exactly did GitHub ship on May 19? GitHub’s May 19 changelog post said Gemini 3.5 Flash is generally available for Copilot and is rolling out to Copilot Pro, Copilot Pro+, Copilot Business and Copilot Enterprise users. The company said the model offers strong tool use, fast response times and high cache efficiency. Google’s model joins a growing list of options inside Copilot rather than replacing the service’s broader multi-model setup. (github.blog) GitHub’s documentation says Copilot supports multiple AI models and that model availability can vary by plan, client and organization settings. ### Why is the “14x premium request multiplier” worth noticing? GitHub said Gemini 3.5 Flash is launching with a 14x premium request multiplier and added that pricing is tentative and subject to change. (github.blog) That matters because GitHub uses model-specific pricing and usage controls across Copilot plans, and the company’s billing documentation now lists per-token pricing and reference rates for additional usage. (docs.github.com) GitHub’s pricing page also says the company will move Copilot from request-based billing to usage-based billing for organizations, enterprises and individuals starting June 1, 2026. That means the cost of choosing among models is becoming more explicit as GitHub adds higher-end and specialized options. ### How does the new desktop app fit into Copilot’s broader push? (github.blog) GitHub said on May 14 that the Copilot app entered technical preview as a desktop experience designed to start “from the work in front of you,” keep tasks isolated and land changes through pull request review. The company described it as a GitHub-native client rather than just another chat window. (docs.github.com) The app is aimed at workflows that begin with GitHub context such as issues, pull requests and repositories. GitHub said users can start from existing work, steer the agent as it progresses and complete the change through normal review flows. ### Is GitHub building toward issue-to-pull-request automation? GitHub has been adding agent features around the pull request lifecycle for months. (github.blog) A March 26 changelog post said users could ask Copilot to resolve merge conflicts on pull requests, while February and April usage-metrics updates tracked pull request outcomes from review suggestions to merged pull requests. Those releases show GitHub tying model choice, coding agents and workflow measurement into the same product surface. In the May 14 post, GitHub said the app can start from GitHub context and “land the change through pull request review,” which places the desktop client inside that broader agent loop. ### Where are these models hosted, and what happens to prompts? (github.blog) GitHub’s hosting documentation says Copilot uses Gemini models hosted on Google Cloud Platform, including Gemini 3.5 Flash. The same page says prompts and metadata are sent to Google Cloud when Gemini models are used, and states that Gemini does not use prompts or responses to train its models. GitHub also says default AI models in Copilot run through its content filters for harmful, offensive or off-topic content, and for public code matching when that setting is enabled. (github.blog) Those controls apply across the supported-model system documented by GitHub. June 1, 2026 is the next concrete date for users to watch. GitHub’s pricing documentation says that is when Copilot moves to usage-based billing, while the Copilot app remains in technical preview and Gemini 3.5 Flash continues rolling out across paid Copilot plans. (docs.github.com 1) (docs.github.com 2) (docs.github.com 3)