Google on Wednesday announced Gemini 4 Carbon, describing it as its next frontier model with stronger coding, long-horizon agent work and cybersecurity scores than Gemini 4 Argon, which the company said it released about two weeks earlier^1,2,3.

The model is not broadly available. Google said it is taking part in the U.S. government’s voluntary pre-release testing process and will start this week with a subset of cyber defenders, then expand gradually through its Fairwind Program before consumer and developer product surfaces^2,1.

Google said Carbon ranks first on Artificial Analysis Intelligence Index version 5.0 with a score of 59, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra on that index^2,1. The company also reported 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, and said those cybersecurity evaluations were run without guardrails².

The Verge led with those cybersecurity claims and the Intelligence Index ranking¹. Ars Technica said Carbon looks like a modest improvement on Argon, that rival frontier models cluster in a similar score range, and that the gaps Google showed are often a few points³.

Google said internal tests showed more than a twofold rise in success generating secure code versus Argon, and that its Chrome team reported Carbon performed more than 10 times better than another lab’s large frontier model at generating high-quality patches^2,3. Wiz, Google’s cloud security platform, said Carbon led on web penetration-testing benchmarks used in its Scan for Good work, according to Google².

On WebDev Arena, Google said Carbon rose 89 Elo points to rank fourth at 1769^2,3. Google also said internal teams have seen the model run for days on multi-step engineering tasks such as root-causing and patching bugs in large codebases^2,1.

On safety, Google said Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, cut unsupported claims by about two times versus Argon, and scored 0.0% attack success on Gray Swan’s Indirect Prompt Injection benchmark, compared with 0.7% for Argon². Ars Technica reported the same 0% Gray Swan figure and compared it with 1% for Opus 5.5 and 8.5% for GPT-6 Astra³.

Google said Carbon will launch at Argon’s introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off the input price^2,1. Ars Technica said that rate is expected to double after the introductory period, that Argon’s discount runs through the end of November, and that Google did not specify an end date for Carbon³.

What remains unknown includes when developers, businesses and consumers will get access, whether Carbon is a new training run or a post-training upgrade of Argon, and how the company-reported scores will hold up outside Google’s charts^3,2.