Google has released Gemini 4 Carbon, its newest frontier AI model, saying it delivers "significant gains" in coding, long-running agentic tasks and cybersecurity defense. In Wednesday's announcement, the company pointed to Carbon's top ranking on the Artificial Analysis Intelligence Index (v5.0) benchmark, ahead of rival models from Anthropic and OpenAI.
The model is an upgrade on Argon, which Google released earlier this month, and for now is available only to a select group of cybersecurity defense companies before it expands to other partners through Google's Fairwind Program. Carbon currently holds the top spot on SEC-bench Pro, a benchmark measuring a model's ability to find real-world software bugs, and on ExploitGym, which measures how well a model can turn a discovered vulnerability into a working exploit.
Built to Grind, Not Just Sprint
Google said Carbon is designed for longer, more autonomous tasks than its predecessors, with internal teams reporting the model can "run for days to persist through dead ends to tackle complex, multi-step engineering tasks." The company also said the model is better at avoiding shortcuts and resisting prompt injection attacks — a persistent vulnerability in AI agents that attackers have used to hijack automated systems — while cutting "unsupported claims" in its outputs by half compared with Argon.
Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, keeping Google's pricing competitive as the major AI labs continue to undercut each other on cost even while racing to claim the top benchmark scores.
The Benchmark Wars Continue
The release lands in the middle of a crowded stretch for frontier AI launches, with Google, Anthropic and OpenAI all having shipped new models within weeks of each other this autumn. Google's decision to debut Carbon first with cybersecurity-focused partners, rather than a broad consumer rollout, suggests the company sees defensive security use cases — finding and patching vulnerabilities before attackers do — as one of the clearest near-term commercial wins for increasingly capable coding models.