Google starts Gemini 4 Carbon rollout with a limited group of cyber defenders

Google will start with selected cyber defenders as it tests safeguards and gathers feedback before a wider release.

Google announced Gemini 4 Carbon on October 28 and said it would start giving access that week to a subset of cyber defenders. The company plans to add trusted partners through its Fairwind Program while it gathers feedback and adjusts safeguards. Google also said it was participating in the U.S. government's voluntary pre-release testing process. The announcement gave no date for general availability across Google's product surfaces.

Carbon follows Gemini 4 Argon, which Google had already launched. The two have different access paths. An organization that already uses Argon should check whether it can join Carbon's initial cohort before planning a deployment. Google announced intended introductory rates of $2 per million input tokens and $10 per million output tokens, with cached input priced at 95% below the regular input rate. Those figures describe the planned launch offer.

Three measurements for three jobs

Google reported 88.2% on SEC-bench Pro for identifying vulnerabilities, 54.8% on ExploitGym for checking exploitability and 71.0% on CWE-bench for producing fixes. These are separate benchmark task sets. They cannot be multiplied together to estimate how many incidents Carbon could resolve from discovery through remediation. Google ran the external evaluations without guardrails, so the reported scores do not directly measure the rollout configuration.

In a security team, a finding must identify affected code and the conditions that expose it. Validation must establish impact in the intended environment. A fix must address the root cause and preserve legitimate behavior. A trial can measure the review burden at each step, including the time spent reproducing findings and the number of proposed fixes that pass existing regression tests.

What longer work changes

Google says its internal teams have seen Carbon pursue engineering problems for days, including subtle bugs in large codebases. That observation points to a use case the early cohort can evaluate. Teams need to decide how long an agent can run, what systems it may inspect and which decisions require a human to review its evidence before it continues. The rollout gives the initial participants a chance to collect those operational measurements before broader access.

[Google's announcement and benchmark references](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)