Gemini 4 Carbon starts with security partners, not a general release

Google has announced Gemini 4 Carbon, but most people cannot choose it in a product yet. The company is beginning with selected cybersecurity defenders and trusted partners, while broader developer, business and consumer access remains undated. For anyone planning an integration, that access boundary is currently more important than the benchmark lead.

The October 28 announcement positions Carbon as a step beyond Gemini 4 Argon in sustained engineering work. Google describes internal runs lasting days, following investigations through failed approaches to find and patch difficult bugs. That is a claim about persistence across a workflow, not just the quality of a single generated answer.

Three tasks that should not collapse into one score

Finding a suspected vulnerability, demonstrating that it is exploitable and repairing its underlying cause are different jobs. Google reports 88.2% on SEC-bench Pro for discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. The company says these external benchmarks were evaluated without guardrails. They therefore should not be read as a promise about the behavior of the final deployed service.

A scanner can produce many plausible warnings while making an engineering team slower if the warnings are expensive to investigate. A patching model can close a test case while changing behavior elsewhere. Carbon’s proposed advantage is to connect these stages; proving that advantage in a particular repository still requires examining accepted fixes, missed defects and regression tests together.

Ars Technica’s reporting also highlights how small some differences between leading models can be. A higher aggregate score is useful evidence, but it does not establish a universal winner for web design, incident response and every programming language.

The introductory price has an endpoint, but not a date

Google quotes $2 per million input tokens and $10 per million output tokens initially, with cached input discounted by 95%. Its footnote lists later prices of $4 and $20 respectively. No expiry date is given for Carbon’s introductory period. A cost comparison that permanently assumes the lower rates would miss half the announced pricing policy.

The immediate practical use of this announcement is to prepare an evaluation, not to promise customers access that does not exist. The AI desk covers models as products as well as benchmarks: a system that can investigate for days also needs a budget, permission boundaries and a way to review what it changed. Carbon will become a meaningful operational choice when that complete package can be tested outside Google’s selected first cohort.