Gemini 4 Carbon

What it is

Gemini 4 Carbon is Google's frontier model, announced as the successor to Gemini 4 Argon, which launched two weeks before Carbon's announcement. It ranks #1 on v5.0 of the Artificial Analysis Intelligence Index at 59. It is pitched at long-horizon agentic engineering: Google says internal teams saw it run for days on tasks such as root-causing and patching subtle bugs in large codebases and low-level systems. On WebDev Arena it gains 89 Elo to reach #4 at 1768.

Cybersecurity results

The three headline cyber benchmarks are SEC-bench Pro (88.2%, finding vulnerabilities), ExploitGym (54.8%, exploitation) and Collinear's CWE-bench (71.0%, patching). Google states these were evaluated without guardrails.

Google also reports vendor-side claims that the page gives no method for:

• More than 2x the success rate of Argon at generating secure code, on internal benchmarks.
• The Chrome team reports Carbon was more than 10x better than another lab's large frontier model at producing high-quality patches.
• Wiz reports Carbon leading its web penetration testing benchmarks.

Safety and alignment figures

• Gray Swan Indirect Prompt Injection benchmark: 0.0% attack success, against 0.7% for Argon.
• Terminal-Bench 4.0: successful cheating 0.0%, attempted cheating 0.4%, with accuracy also improving.
• High-severity misalignment alerts in internal live coding dropped by almost 80%.
• Unsupported claims are cut by about 2x versus Argon.
• Overreach is down 26% and under-service down 29%.

Access and pricing

Carbon is not generally available. It goes first to a subset of cyber defenders, then to more trusted partners in phases through the Fairwind Program. Google says it is taking part in the U.S. government's voluntary pre-release testing process. It gives no date for availability to developers, businesses or consumers, saying only "as soon as possible".

Pricing is the introductory rate Argon had: $2 per million input tokens and $10 per million output tokens. Cached input tokens cost 95% less than the input price.

What to do

Nothing can be deployed yet. If you do defensive security work, the route in is the Fairwind Program. Treat the numbers as Google's own: the page links no methodology, and the secure-code, misalignment, honesty and overreach figures come from internal testing.

Caveats

• The cyber scores were measured without guardrails, so shipped behaviour may differ.
• Google says it is still iterating on guardrails before wider release.
• The page does not describe the Fairwind eligibility criteria or application process.
• The pricing is labelled introductory, and the page does not say how long it lasts.