Google releases Gemini 4 Carbon, a cybersecurity-focused frontier model

A security team benchmarked a model three weeks ago and signed a contract on the result. Google has now released two frontier models since then. On 28 October, Google announced Gemini 4 Carbon. The company says it is its strongest model. It focuses on cybersecurity and long-horizon engineering work. It arrives two weeks after Gemini 4 Argon.

Google says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. The company reports three cybersecurity benchmark results, all its own. SEC-bench Pro for finding vulnerabilities: 88.2 percent. ExploitGym for exploitation: 54.8 percent. CWE-bench for patching: 71.0 percent. Google also says Carbon climbs 89 Elo points on WebDev Arena to rank fourth, at 1768.

Who gets access first

Carbon is not open to developers yet. Google says it is enrolled in the US government's voluntary pre-release testing process. From this week, the model goes to a group of cyber defenders through what Google calls the Fairwind Program. Google says it will phase in more trusted partners while it gathers feedback on guardrails. A broader launch to developers, businesses, and consumers will follow, but Google does not give a date.

Pricing and the step after the trial

Pricing starts at the Argon rate. The input price is 2 dollars per million tokens. The output price is 10 dollars per million tokens. Cached input tokens are discounted 95 percent. Google states one footnote plainly. After the introductory period, input rises to 4 dollars per million tokens. Output rises to 20 dollars per million. A team pricing a Carbon integration should model both numbers, not only the first.

The security claims Google makes about itself

Google says Carbon doubles the success rate on secure code generation over Argon on internal benchmarks. The Chrome team saw Carbon generate high-quality patches at more than 10 times the rate of another lab's large model. Google states the comparison but does not name the other model. Wiz is Google's named cloud security partner. Wiz reports Carbon leading on web penetration testing benchmarks the company runs. All of these figures come from Google or from a Google partner.

The model also handles agentic work. Google says internal tasks ran for days. The company reports internal results on model alignment too. High-severity misalignment alerts dropped 80 percent against Argon. Unsupported claims dropped by half. Attack success on the Gray Swan prompt-injection benchmark fell to 0 percent, from 0.7 percent on Argon.

The pace is the context. Argon went general availability on 13 October. Carbon followed on 28 October. A buyer of an AI product should ask the vendor two questions. Which Gemini model is it built on? When was the vendor's own test last re-run against the live model? Google's security figures all come from Google. A team worried about the vendor evaluating itself should build its own benchmark on files it owns. Re-run that benchmark on every model that reaches the team through the vendor.

Source

• Introducing Gemini 4 Carbon, Google