NewsAI

Google unveils Gemini 4 Carbon, its strongest model for cybersecurity and long-horizon tasks

Google announced its next frontier model, Gemini 4 Carbon, on October 28, positioning it as the company’s strongest model with state-of-the-art cybersecurity and long-horizon capabilities. Carbon tops the Artificial Analysis Intelligence Index v5.0 with a score of 59 and will be rolled out to cyber defenders this week through the Fairwind Program, priced the same as Gemini 4 Argon, launched two weeks ago: $2 per million input tokens and $10 per million output tokens.

Key takeaways

  • Ranks #1 on Artificial Analysis Intelligence Index v5.0 at 59
  • Leads three external cybersecurity benchmarks: SEC-bench Pro 88.2% (vulnerability finding), ExploitGym 54.8% (exploitation), CWE-bench 71.0% (patching)
  • Jumps +89 elo to #4 (1769) on WebDev Arena
  • More than 2x uplift over Argon in secure code generation success rate
  • Cheating on Terminal-Bench 4.0 reduced to 0.0%; 100% resistance on Gray Swan prompt injection benchmark
  • Priced same as Argon: $2/$10 per million tokens, cached input at 95% off

Leading the full cybersecurity defense lifecycle

Carbon’s headline strength is the cybersecurity defense lifecycle: finding vulnerabilities, validating them, and fixing them. Evaluated without guardrails on external benchmarks, the model achieves a state-of-the-art 88.2% on SEC-bench Pro, autonomously uncovering complex flaws buried in real-world software, a new record 54.8% on the exploitation benchmark ExploitGym, and 71.0% on Collinear’s CWE-bench for root-cause patching, turning validated findings directly into working fixes. Internal numbers are equally aggressive: more than 2x uplift over Argon in secure code generation success rate, a report from the Chrome team that Carbon performed more than 10X better than another lab’s large frontier model at generating high-quality patches, and a Wiz observation that Carbon leads web penetration testing benchmarks. Google also says Carbon has been seen running for days internally, persisting through dead ends on multi-step engineering tasks such as root-causing subtle bugs across large codebases.

Alignment gains: zero cheating, 100% injection resistance

Safety is where Carbon most clearly separates itself from Argon. Building on the same multi-layered stack, model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring, the model reduced successful cheating on Terminal-Bench 4.0 to 0.0% (attempts down to 0.4%) while improving overall accuracy, cut high-severity misalignment alerts in internal live coding environments by almost 80%, and reduced unsupported claims by roughly 2x over Argon. Overreach fell by 26% and under-service by 29%, and the model now surfaces discrepancies to users rather than guessing when facing ambiguity. On Gray Swan’s Indirect Prompt Injection benchmark, Carbon pushed Argon’s already-low 0.7% attack success rate to 0.0%, resisting every evaluated attack. The rollout is deliberately staged: Google is engaged in the U.S. government’s voluntary pre-release testing process, and access starts this week with a subset of cyber defenders via the Fairwind Program before broader availability to developers, businesses and consumers.

Source: Google Blog (2026-10-28)