Gemini 4 Carbon puts Google’s cybersecurity claims behind a restricted rollout

Google announced Gemini 4 Carbon on October 28 as its newest frontier model, emphasizing cybersecurity and long-running engineering work. The company reports scores of 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. It also says Carbon reached fourth place on WebDev Arena with a score of 1,768, up 89 Elo points. Those figures describe distinct evaluations, not a single end-to-end measure of how well the model secures production software.

The timing matters: Carbon follows Gemini 4 Argon by about two weeks, according to Google. Google frames Carbon as a step forward in coding, agentic tasks, scientific work, web development, and multimodal understanding. Its announcement says internal teams saw the model persist for days on difficult engineering tasks, including root-causing and patching subtle bugs across large codebases. These are Google’s reported internal observations, not independently documented results from customer deployments.

Google’s most concrete comparison with Argon is its claim that Carbon more than doubles the success rate for generating secure code on internal benchmarks. It also reports lower unsupported claims and fewer high-severity misalignment alerts in internal coding environments. Reuters independently reported the model launch, its focus on cybersecurity, the claimed improvement in secure-code generation, and the comparison with Argon. The detailed benchmark scores and internal safety measures remain Google-reported; the cited source set does not independently reproduce them.

Access is the main practical limit. Google says it is starting with a subset of cyber defenders and will gradually add trusted partners through Fairwind while it gathers feedback and iterates on guardrails. The company says broader access for developers, businesses, and consumers will follow, but gives no date. Reuters also describes the initial defender-only availability. For teams outside the program, Carbon’s reported scores are not yet a basis for direct testing or migration decisions.

Google lists introductory prices of $2 per million input tokens and $10 per million output tokens, matching Argon, with cached input tokens discounted by 95%. The announcement says those prices later rise to $4 and $20 per million tokens, respectively. That gives developers a useful cost reference, but not a full comparison of task economics: the public announcement does not provide independently verified evaluation runs, workload-level costs, or general access. For now, Carbon is a signal of Google’s intended direction in security-focused models, rather than an option most teams can assess hands-on.