Google DeepMind announced Gemini 4 Carbon, reporting gains on long-context, knowledge-work automation, and science and math benchmarks.
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it the company’s strongest model and highlighting long-horizon work, science and cybersecurity. Google says it will initially provide access to a subset of cyber defenders, then expand to trusted partners through its Fairwind program; it has not announced a date for general availability. The staged rollout reflects the model’s dual-use capabilities: defenders may use it to find and fix software flaws, while broader access is being controlled. Google reports 96.2% on GraphWalks’ 256,000-to-1-million-token long-context test, versus 84.2% for Gemini 4 Argon and 71.8% for GPT-6 Astra. Carbon scored 53.9% on AutomationBench, 80.0% on RiemannBench and 64.8% on Terminal-Bench-Science. In security evaluations, it reached 88.2% on SEC-bench Pro for vulnerability discovery and 71.0% on CWE-bench for patching. These are company-reported benchmarks, not independent proof of real-world performance.