Gemini 4 Carbon, the newest frontier model from Google, ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Access starts this week for a subset of cyber defenders. Additional trusted partners follow in phases. Developers, businesses and consumers get it through Google's products later, as Google gathers feedback and iterates on the guardrails.
Google calls the rollout deliberate. It reports Carbon at 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching, all run without guardrails. Gemini 4 Argon, which reached trusted defenders at the end of September, tied for first on CWE-bench v1 at 68%. Collinear's v1 leaderboard lists both models, with Carbon three points ahead. On secure-code generation, Google says its internal benchmarks show Carbon succeeding more than twice as often as Argon. Google's Chrome team also reported that Carbon produced high-quality patches more than 10 times better than another lab's frontier model.
The additional trusted partners come through the Fairwind Program, which gives high-priority defenders such as governments, healthcare providers and telecommunications services early access to models that find and fix vulnerabilities. DeepMind says it works with more than 650 partners globally. Google is also taking part in the U.S. government's voluntary process for pre-release model testing.
Pricing stays where Argon set it. Carbon will launch at the same introductory rate, and Google's footnote puts the rate after that period at double. Google does not say when the introductory period ends.
| Per million tokens | Introductory | After introductory period |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cached input | 95% off input price (about $0.10) | Not stated |
Outside security, Google says Carbon is built for long engineering jobs. Internal teams have watched it run for days on tasks such as tracing and patching subtle bugs across large codebases and low-level systems. On WebDev Arena it gained 89 Elo points and sits fourth at 1768.
Google also reports alignment results. Carbon's successful cheating rate on Terminal-Bench 4.0 is 0.0%, with cheating attempts at 0.4%. High-severity misalignment alerts in live coding sessions with Google's teams dropped by almost 80%. On Gray Swan's indirect prompt injection benchmark, where Argon let through 0.7% of attacks, Carbon resisted every evaluated attempt.
