Gemini 4 Carbon Aims for Cyber Defense with Record Scores

The new Gemini 4 Carbon frontier model is introduced with advancements in cybersecurity, alignment, and long engineering tasks. The publisher reports a 0.0% success rate on evaluated prompt injection, a top ranking in an external index, and gains across multiple benchmarks, while announcing a gradual rollout and a price identical to Argon.

Alignment and operational security reported to be significantly improved

The publisher claims that Gemini 4 Carbon reduces the successful cheating rate to 0.0% on Terminal-Bench 4.0, with attempts limited to 0.4%, while improving overall accuracy. In live coding, high-severity misalignment alerts reportedly dropped by nearly 80%. Internal tests also indicate about half as many unsupported claims compared to Argon, with the integration of methodological warnings directly into deliverables or the postponement of actions until required elements are available. The model is described as consulting relevant sources and exposing potential discrepancies to the user before taking action. The publisher also reports a simultaneous decrease of 26% in overreach and 29% in insufficient service, leading to more reliable execution of multi-step plans within the defined framework. On the Gray Swan benchmark for indirect prompt injection, Argon showed a 0.7% attack success rate, while Carbon is credited with 0.0%, resisting all evaluated attempts. These promises are part of a safety stack shared with Argon, combining model training, real-time anti-abuse defenses, resilience to injections, and thought chain monitoring, with specific training for complex or high-pressure tasks. The publisher also mentions previously documented alignment improvements for Argon.

Cybersecurity: detection, exploitation, and remediation reported at the forefront

Gemini 4 Carbon is presented as making progress across the entire defense cycle: detection, risk validation through exploitation, and remediation. The publisher reports 88.2% on SEC-bench Pro for detecting real vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on the CWE-bench for root patching, noting that these evaluations were conducted without safeguards. Internally, the success rate for generating secure code is said to have more than doubled compared to Argon. The Chrome team reports that Carbon performed over 10 times better than another frontier model in generating high-quality patches. On the partner side, Wiz claims that the model dominates its web penetration testing benchmarks, a capability highlighted as essential to its Scan for Good initiative. The publisher characterizes SEC-bench Pro as focused on real vulnerabilities, ExploitGym as a popular exploitation benchmark, and the CWE-bench from Collinear as oriented towards root cause patching.

Long tasks and web development: claimed gains

The model is designed for complex and sustained engineering challenges, with supported agentic workflows. Internally, the publisher indicates that Carbon can operate for days to unlock situations and handle multi-step tasks, such as identifying root causes and correcting subtle bugs in large codebases and low-level systems. Efforts have reportedly targeted improvements in coding, agentic and scientific tasks, full-stack web development, and multimodal understanding. On WebDev Arena, the publisher claims a jump of +89 Elo points, a #4 rank, and a score of 1769, with increased output and artifact quality. Internal benchmarks report improvements in secure code generation and live coding environments with reduced manufacturing rates.

Controlled access: voluntary testing and Fairwind program

The publisher states it is participating in the voluntary model testing process of the U.S. government before launch. Starting this week, access to Gemini 4 Carbon is set to begin for a subset of cybersecurity defenders, followed by a gradual extension to trusted partners through the Fairwind program. The collection of feedback and iteration on safeguards is presented as ongoing before making it available to developers, businesses, and consumers through the publisher's product surfaces when possible. The company indicates it is awaiting usage by the first group of defenders and aims for broader availability soon.

Unchanged pricing and positioning against Argon

The announced introductory price remains the same as for Argon: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. Gemini 4 Carbon builds on Gemini 4 Argon, which was launched two weeks ago. The publisher claims to be rapidly advancing the frontier, having its teams and models work in concert while simultaneously strengthening protections.

External ranking and synthetic indicators

The publisher highlights a first place for Gemini 4 Carbon in the recently updated v5.0 version of the Artificial Analysis Intelligence Index, with a score of 59. It reports gains in coding and agentic tasks and reminds that some cited external benchmarks were conducted without safeguards. The stated goal remains to bring advancements in software engineering and cyber defense without sacrificing security or reliability.