Michael JohnsonGoogle announced Gemini 4 Carbon on October 28, 2026, presenting it as a new system for long-running engineering tasks and cybersecurity. The company plans to give an initial group of defenders access first, then expand access through the Fairwind program. Google says other developers, businesses, and consumers will eventually be able to access it through its products, but has not given a specific date.
The headline figures concern specific security tasks. According to Google’s announcement, Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for assessing exploitability, and 71.0% on CWE-bench for preparing fixes. Google describes these as leading results on the respective benchmarks and notes that the external evaluations were conducted without safeguards.
From finding a flaw to fixing it
The set of metrics covers several stages of a defender’s work: identifying a potential weakness in real software, checking its practical significance, and proposing a fix. This breakdown can help security teams choose tasks for a pilot: they can separately evaluate discovery quality, risk confirmation, and the usability of a generated patch, then put it through their usual review and testing.
Google also says that, in its internal tests, Carbon more than doubled the likelihood of generating secure code compared with Gemini 4 Argon. This is the company’s result from its own tests and conditions.
Access and pricing do not yet amount to a general launch
According to Google’s plan, some cybersecurity professionals will get the model at the first stage, after which the company intends to bring on additional trusted partners. Google said it is gathering feedback and will continue refining its safeguards before expanding access. Access for developers, businesses, and consumers through Google products is planned for the future, with no specific timeline.
The company announced introductory pricing of $2 per million input tokens and $10 per million output tokens; cached input tokens cost 95% less than regular input tokens. After the introductory period, the stated rates are $4 and $20, respectively. These prices offer a planning reference, but most teams will be able to assess costs in practice only after getting access and testing the model on their own workloads.
An earlier article about Google DeepMind’s Genie 4 covered interactive simulations; Carbon is a separate announcement addressing a different task. For security professionals, the immediate practical question is whether they can join the first testing group. Others should rely on the published terms and metrics until Google announces specific access dates.
