Gemini 4 Carbon reaches GitHub Copilot, but only on paid Pro+ and above

GitHub, owned by Microsoft, said that Google’s Gemini 4 Carbon is now generally available in GitHub Copilot. According to the changelog post, it is available on Copilot Pro+, Max, Business and Enterprise plans, and it is billed at provider list pricing.

Paid plans only, rolled out gradually

GitHub describes the model as built for complex coding, agentic and reasoning work, with an emphasis on secure code generation and security-focused tasks. Access is limited to Pro+, Max, Business and Enterprise users. The post does not say whether Copilot Free or standard Pro accounts will get it.

Users can choose it in the model picker across a wide range of surfaces: Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode and Eclipse. GitHub says the rollout is gradual, so some accounts may not see the model yet.

The model is billed under usage-based billing at what GitHub calls provider list pricing, and it points readers to its Models and pricing page for the figures. The changelog itself gives no rates, no multipliers and no context-window size.

For organizations, Business and Enterprise administrators control access through the model policy in Copilot settings. Under default model enablement, new models switch on automatically, unless an administrator has turned off the global default or disabled this particular model. Teams that want a review step before developers can use Carbon therefore need to act on the policy rather than wait.

List pricing makes the Gemini API rate the benchmark

Because Copilot bills at provider list pricing, Google’s own price is the reference point. According to our earlier report on Cognition’s test, Google lists Carbon in the Gemini API at $2.00 per million input tokens and $10.00 per million output tokens. Those introductory rates run through November 30, 2026, and the post-promotion price was not stated. Whether Copilot passes through the introductory rate is not established by GitHub’s post, so teams should check the pricing page before switching heavy agent workloads over.

Usage-based billing also changes the calculation for agentic work. A coding agent that reads a repository, plans, edits and runs tests can consume far more tokens than a chat answer, and the bill scales with that. Cognition said Carbon’s best result costs under two-thirds as much per task as the best result of Gemini 4 Argon, though it did not give dollar figures, and that comparison was run in Devin rather than Copilot.

One vendor benchmark, and none for security

GitHub’s post makes no benchmark claims. The most concrete outside number comes from Cognition, the company behind Devin, which ran its FrontierCode 2.0 test. In our coverage, Carbon scored 40.0%, up from 26.3% for Argon, and close to GPT-6.1 Sol at 39.0%. It still trailed Claude Fable 5.2 at 49.1% and Claude Opus 5.5 at 45.5%. Cognition built and runs that benchmark, and the results have not been independently reproduced.

The security emphasis in the GitHub post deserves the same caution. Google described Carbon as strong at cybersecurity tasks when it unveiled the model on October 28, and it first reached a limited group of defenders. Cognition’s separate security test, covered in our Mythos 5.2 report, used 25 open-source repositories with 1,781 known bugs, but Carbon was not part of it. So there is no public figure here showing how Carbon performs at finding or avoiding vulnerabilities in code, and GitHub’s post offers none.

What the post does establish is distribution. Carbon sits next to other models in one picker, which lets developers compare outputs on their own repositories instead of relying on vendor scores. For teams already on a qualifying plan, that is the practical test: run the same tasks with Carbon and the current default, then compare quality and cost.

The same week, GitHub also changed how enterprises administer Copilot, with seat management that can run from a GitHub App rather than a person’s token, as we reported. The two items are separate, but both point to Copilot being managed as shared infrastructure, with model access governed by policy.