Google DeepMind officially released its next-generation frontier model, Gemini 4 Carbon, today. Google describes it as the most capable model to date, emphasizing cybersecurity defense and long-horizon task capabilities. This launch comes just two weeks after the release of the previous generation, Gemini 4 Argon.
Carbon scored 59 on the updated Artificial Analysis Intelligence Index v5.0, ranking first. Google reported three cybersecurity benchmark results: 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitability verification, and 71.0% on CWE-bench for root cause remediation, all achieving state-of-the-art performance. On WebDev Arena, its score improved by 89 Elo points to 1768, placing it fourth.
Google's internal tests show that Carbon's success rate in generating secure code is more than double that of Argon. The Chrome team stated that Carbon's performance in generating high-quality patches is over ten times better than that of another lab's large frontier model; meanwhile, cloud security platform Wiz noted its lead in web penetration testing benchmarks. Carbon is designed for long-running agent workflows, and Google internally ran it continuously for several days to locate and fix subtle defects in large codebases and low-level systems.
Carbon retains Argon's multi-layered protection, which includes model training, real-time abuse defense, anti-prompt injection, and chain-of-thought monitoring. Google stated that Carbon's successful cheating rate on Terminal-Bench 4.0 dropped to 0%, with cheating attempts down to 0.4%. The company observed an nearly 80% reduction in internal high-risk alignment alerts, and unsubstantiated conclusions were cut by about half compared to Argon. In Gray Swan's indirect prompt injection benchmark, the attack success rate fell from Argon's 0.7% to 0%. Carbon also saw a 26% reduction in over-intervention and a 29% reduction in under-service.
Google is rolling out Carbon in phases: starting this week with a small group of cybersecurity defenders, then gradually expanding through the Fairwind Program to developers, enterprises, and consumers, though no specific timeline has been provided. Carbon's pricing follows Argon's introductory rates: $2 per million input tokens and $10 per million output tokens, with cached inputs billed at a 95% discount off the input price. After the introductory period, prices will adjust to $4 and $20 respectively.
