Google announces limited cyber-defense rollout for Gemini 4 Carbon

Google announced Gemini 4 Carbon on October 28 and said access to the new frontier model would begin this week for a subset of cyber defenders. Google plans to expand access gradually to additional trusted partners through the Fairwind Program.

For now, developers, businesses and consumers cannot generally use Carbon. Google said it is gathering feedback and refining guardrails before bringing the model to its product surfaces, but gave no general-availability date and did not identify the initial users. It also left the cohort’s size, selection criteria, geographic scope and permitted uses undisclosed.

Google presented Carbon as an upgrade for coding, long-running agentic work and cybersecurity. The company said internal teams have run it for days on multi-step engineering tasks, from tracing root causes to patching bugs across large codebases and low-level systems. The evidence available for this story does not include an independent evaluation of those real-world performance claims.

Google also published several benchmark figures. Artificial Analysis independently lists Carbon (High) first with a score of 59 on version 5.0 of its Intelligence Index. Google reported 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on Collinear’s CWE-bench for patching. CWE-bench’s public leaderboard lists Carbon at 71%. Google said the security evaluations were run without guardrails. The SEC-bench Pro and ExploitGym scores, and Google’s state-of-the-art characterizations, were not independently reproduced in the opened evidence.

Google also cited performance figures from its own teams and Wiz. Google claimed more than a twofold improvement over Gemini 4 Argon on internal secure-code benchmarks. It said its Chrome team achieved more than tenfold better patch generation than an unnamed lab’s frontier model and that Wiz observed Carbon leading web-penetration-testing benchmarks. The announcement did not provide the methods or underlying results needed to assess those comparisons independently.

For alignment and safety, Google reported zero successful cheating and a 0.4% cheating-attempt rate on Terminal-Bench 4.0, an almost 80% reduction in high-severity misalignment alerts in internal coding environments, and a 0.0% attack-success rate on Gray Swan’s indirect prompt-injection benchmark. Those results remain Google claims. The company said Carbon uses the same multilayer safety stack as Gemini 4 Argon, including training, real-time misuse defenses, prompt-injection protections and chain-of-thought monitoring. It did not publish enough implementation detail for an outside assessment.

Google also said it is participating in the U.S. government’s voluntary process for pre-release model testing. The available evidence did not include confirmation from the relevant government body.

Google said Carbon will initially cost $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached inputs. Google said the rates will rise to $4 and $20 after an unspecified introductory period. The announcement did not establish when general paid access will begin or whether the introductory prices apply during the controlled rollout.