Google introduced Gemini 4 Carbon with an initial cybersecurity-focused rollout. Its performance figures are company-reported benchmarks, and access is limited.
Google has introduced Gemini 4 Carbon, a new model it is positioning around demanding cybersecurity work and longer, multi-step tasks. The announcement, published on October 28, describes an initial rollout aimed at selected cyber defenders, followed by access for trusted partners through the company’s Fairwind programme.
This is not a general public release in the usual sense. Google’s first phase is deliberately constrained because a capable model can help defenders identify vulnerabilities and analyse systems, but the same capabilities require careful controls. The staged approach is intended to test the model with organisations that can apply security processes and responsible-use safeguards.
What Google says makes it different
Google highlights performance on its internal and third-party evaluation set, including Artificial Analysis and cybersecurity-oriented benchmarks. The company reports an Artificial Analysis Intelligence Index score of 59 and describes strong performance on selected security tasks. These are Google’s reported results; they should be compared with independent evaluations and real-world performance as more testing becomes available.
The company also lists introductory API pricing of $2 per million input tokens and $10 per million output tokens. Pricing and access terms can change, and the limited rollout means availability may matter more than the headline price for most developers today.
Why defenders should pay attention
Security teams spend time reading logs, triaging reports, tracing dependencies and testing potential vulnerabilities. AI can accelerate parts of that workflow, but it cannot replace permission boundaries, human review or responsible disclosure. The key question is whether an organisation can safely verify, prioritise and fix what a model finds.
For now, Gemini 4 Carbon is best understood as a specialised, controlled launch. Wider access and independent testing will show whether its benchmark claims translate into a dependable everyday tool.
