Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon Model

What it claims

Gemini 4 Carbon is a Google frontier model that, by Google's own comparison, leads GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on vulnerability detection (SEC-bench Pro, 88.2%), exploitation (ExploitGym, 54.8%) and patching (Collinear's CWE-bench, 71.0%). It follows Gemini 4 Argon, launched two weeks earlier. Google presents exploitation as a defensive task. Its stated reason is that defenders use models to check which vulnerabilities are exploitable before deciding what to remediate first, or at all.

Evidence

All headline scores were run with no guardrails in place, so they describe the raw model, not the version that will ship. The rival scores are not given on the page, only that Carbon tops them.

Further claims, mostly against Argon:

• Secure code: more than twice as successful as Argon, on Google's internal benchmarks.
• Patch quality: Google's Chrome team found Carbon beat "a large frontier model from a rival lab" by more than 10 times. The rival is not named and the metric is not given.
• Web pentesting: Wiz, which Google owns, reports Carbon in the top spot on web penetration testing benchmarks. The benchmarks are not named.
• Prompt injection: no attack succeeded in Gray Swan's indirect prompt injection benchmark, against a 0.7% success rate for Argon.
• Benchmark cheating: on Terminal-Bench 4.0, none of Carbon's cheating attempts succeeded, and attempts fell to 0.4%. No baseline is given for the earlier rate.
• Reliability: about half as many unsupported claims as Argon.
• Misalignment: internal teams using Carbon in real coding environments saw a nearly 80% decrease in high-severity misalignment alerts.

How to use it

General developers cannot use Carbon yet. Access opens first to a select group of cyber defenders, then widens in stages to approved partners through Google's Fairwind Program. Fairwind gives governments, critical infrastructure operators and major technology platforms early access to models for cyber defense. Google says more than 650 organizations take part.

Pricing matches Argon's introductory rates: $2 per million input tokens and $10 per million output tokens. When the introductory period ends, the rates rise to $4 and $20. The page does not say when that period ends. It gives no API model identifier and no context window.

Limits

Every result comes from Google or Google-owned Wiz, and none of it is independently verified. The comparisons with rival models carry no per-model scores. Google says it is still collecting feedback and iterating on guardrails before a release to developers, businesses and consumers. It gives no date for that release. Google also says it takes part in the US government's voluntary pre-release model testing effort. The page does not state the results of that testing.