Google DeepMind announces Gemini 4 Carbon

Google DeepMind announced Gemini 4 Carbon with a cybersecurity focus and called it its next frontier model.[1][2] The company also described Carbon as state-of-the-art in cybersecurity and long-horizon capabilities.[2] Google called Carbon its strongest model.[3] These are company characterizations; the available information does not include independent evaluations or benchmark results that would let teams compare those claims. For professionals assessing the announcement, the distinction matters: the claims indicate the intended emphasis, but do not establish how Carbon performs in a particular organization’s environment.

Google DeepMind says Carbon can work through complex tasks for days without getting stuck or taking shortcuts.[2] The company also says the model delivers high-quality, reliable code.[2] The announcement does not provide task examples, performance measurements, or information about how these qualities were assessed. Teams considering coding or extended-task uses should treat these as claims to validate against their own requirements, rather than as evidence that the model will complete a given workflow without review.

Google DeepMind says it has received feedback from an initial group of cyber defenders and that the feedback will help it iterate on guardrails.[2] The company says it plans to expand access soon.[2] The evidence does not specify who will be eligible, when access will open, or what safeguards will be in place. That leaves practical deployment questions unresolved for teams weighing evaluation plans, including the access route and how guardrails will apply to their intended work.

It also reported that Carbon had been deployed on Google’s internal coding platform, Jetski, over the preceding few days, and that Carbon was said to beat Argon mainly on programming tasks.[4] The Decoder reported that one employee compared Carbon with Anthropic’s Opus 5.5 and said it needed more testing.[4] It also reported that an employee found early Argon versions reminiscent of older Opus 5 on some coding tasks, while internal feedback on Argon was positive overall.[4] These reports describe internal impressions, not published comparative test results.