Google Holds Back Gemini 4 Carbon, Gives Cyber Defenders First Access After Reported 88.2% Vulnerability-Finding Score
Google Holds Back Gemini 4 Carbon, Gives Cyber Defenders First Access After Reported 88.2% Vulnerability-Finding Score
Google DeepMind has introduced Gemini 4 Carbon, positioned by the company as its most capable model to date and the successor to Gemini 4 Argon. The announcement places particular emphasis on cybersecurity benchmark performance, long-horizon agentic coding capability, and a set of safety improvements over the prior model. It's worth noting upfront that these claims originate entirely from Google's own announcement and associated coverage, and have not yet been independently verified by outside researchers or benchmarking organizations.
Google DeepMind Unveils Gemini 4 Carbon
According to Google, Gemini 4 Carbon represents a meaningful step forward from Gemini 4 Argon, with particular attention paid to how the model performs on cybersecurity-focused benchmarks and extended, multi-step agentic coding tasks. Many observers note that the framing of the release leans heavily on benchmark scores and safety metrics reported by Google itself, a pattern common to major model launches but one that invites some caution before treating the figures as settled fact.
Why Cyber Defenders Get First Access
Rather than a broad public release, Google is reportedly pursuing a phased rollout that gives cybersecurity defenders and what the company calls "trusted partners" early access to the model. The stated rationale centers on the dual-use nature of a system that is unusually effective at finding vulnerabilities: the same capability that helps defenders patch weaknesses could, in principle, be used to locate and exploit them. Google frames this staged approach as a responsible-deployment measure. A recurring consumer and industry concern, however, is distinguishing genuine safety-driven caution from a rollout strategy that also serves marketing and exclusivity purposes, and the available coverage does not provide enough independent detail to fully separate the two motivations.
The Headline Number: 88.2% on SEC-bench Pro
Much of the coverage centers on a reported score of 88.2% on SEC-bench Pro, a benchmark described as measuring a model's ability to identify security vulnerabilities. Google has also cited related figures, including 54.8% on ExploitGym and 71% on CWE-bench, as further evidence of the model's security-oriented capabilities. For broader context, the company says Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. These numbers are presented by Google as straightforward achievements, but as with other vendor-reported benchmarks, they should be treated as provisional until independent testing organizations have had the opportunity to reproduce or scrutinize them.
Coding, Agentic Workflows, and Safety Claims
Google reports that Gemini 4 Carbon more than doubles Argon's success rate on internal secure-code generation benchmarks, and the company's Chrome team is said to have observed roughly a tenfold improvement in patch generation compared with another frontier model. The model is also credited with an 89 Elo point gain on WebDev Arena, placing it fourth overall at a rating of 1,768.
On the safety side, Google's self-reported metrics include a nearly 80% reduction in high-severity misalignment alerts and roughly a 50% drop in unsupported claims generated by the model. The company also reports a 26% reduction in what it terms model overreach and a 29% reduction in under-service, along with a 0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark. On Terminal-Bench 4.0, Google reports a 0% rate of successful cheating behavior and a 0.4% rate of attempted cheating. These figures are notable, but it's important to emphasize that they come directly from Google rather than from an independent auditor, and no third-party confirmation of these specific safety metrics is currently available in the sourcing reviewed for this article.
Pricing and Availability
Google has disclosed pricing for Gemini 4 Carbon at $2 per million input tokens and $10 per million output tokens, with a 95% discount applied to cached input tokens. For now, access remains limited to cybersecurity defenders and trusted partners, with Google indicating that broader availability will follow, though no firm timeline has been specified in the available coverage.
A Pattern of Unverified Claims
A recurring theme across coverage of Gemini 4 Carbon is the extent to which secondary reporting closely mirrors the language and figures of Google's original announcement. Many outlets appear to have republished the company's benchmark numbers and safety statistics with little independent scrutiny or additional verification. This doesn't necessarily indicate that the claims are inaccurate, but it does mean that, at this stage, the performance and safety figures associated with Gemini 4 Carbon should be understood as vendor-reported rather than independently confirmed. Readers interested in a fuller picture may want to watch for third-party benchmarking or academic analysis as the model becomes more widely available, particularly given its dual-use relevance to cybersecurity.