Gemini 4 Carbon

What it claims

Gemini 4 Carbon is Google's new frontier model, aimed at long-horizon agentic coding and cybersecurity defense. It has no general availability yet. Google says Carbon ranks #1 on Artificial Analysis Intelligence Index v5.0 (score 59) and leads SEC-bench Pro (88.2%, vulnerability finding), ExploitGym (54.8%, exploitation) and Collinear's CWE-bench (71.0%, patching). The page does not give the announcement date. It says Gemini 4 Argon launched two weeks earlier.

Evidence

The three cybersecurity scores were measured without guardrails. The deployed model may score differently. The page gives no competitor scores for these benchmarks, only the "state-of-the-art" label.

Other results the page reports:

• WebDev Arena: Carbon is #4 at 1768 Elo, up 89 points. The page does not name the baseline, but Argon is the implied one.
• Secure code generation: Carbon has more than 2x the success rate of Argon on internal benchmarks.
• Patch quality: The Chrome team reports Carbon did more than 10x better than "another lab's large frontier model". That model is not named.
• Terminal-Bench 4.0: Successful cheating fell to 0.0% and cheating attempts to 0.4%, while accuracy rose.
• Gray Swan Indirect Prompt Injection Benchmark: The attack success rate is 0.0%. Argon's was 0.7%.
• Internal coding environments: High-severity misalignment alerts fell by almost 80%. Unsupported claims fell about 2x against Argon. Overreach fell 26% and under-service fell 29%.

Wiz reports that Carbon leads its web penetration testing benchmarks. The page gives no figures for this.

How to use it

Access starts with a subset of cyber defenders and will widen to trusted partners through the Fairwind Program. Developer, business and consumer access is promised "as soon as possible," with no date. The page names no API model ID and no endpoint.

Introductory pricing matches Argon:

• $2 per million input tokens.
• $10 per million output tokens.
• Cached input at 95% off the input price, which works out to $0.10 per million.

Limits

• Most alignment and coding-quality figures come from internal benchmarks or Google's own teams, with no setup described.
• The cyber benchmarks were run without guardrails, so production behavior may differ.
• Claims that Carbon runs "for days" on engineering tasks are anecdotal and unquantified.
• Google is taking part in the U.S. government's voluntary pre-release testing process. It is still iterating on guardrails before broad release, so behavior and access terms may change.
• The page says the price is introductory but does not say how long it will last.