Gemini 4 Carbon is our strongest model, featuring state‑of‑the‑art cybersecurity and long‑horizon capabilities. Carbon ranks #1 on the latest v5.0 Artificial Analysis Intelligence Index (59) and shows significant gains in coding, agentic tasks, and key security abilities. It scores 88.2% on SEC‑bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE‑bench for patching. Internal benchmarks report more than a 2× uplift in secure code generation and an ~80% drop in misalignment and fabrication rates in live coding.
Built on Gemini 4 Argon released two weeks earlier, Carbon illustrates the rapid co‑evolution of models and research. We are participating in the U.S. government’s voluntary pre‑release testing and this week are granting access to a subset of cyber defenders, followed by a gradual rollout to trusted partners via the Fairwind Program, and eventually to developers, businesses, and consumers. Pricing matches Argon: $2 per million input tokens, $10 per million output tokens, with cached input tokens at a 95% discount.
Carbon is designed for sustained agentic workflows, running for days to tackle multi‑step engineering problems such as root‑causing and patching subtle bugs across large codebases and low‑level systems. Full‑stack web development and multimodal understanding also improve, lifting WebDev Arena by 89 elo points to #4 (1768). Across the cybersecurity lifecycle—detection, validation, remediation—Carbon achieves state‑of‑the‑art results.
On safety and alignment, Carbon inherits Argon’s multi‑layer stack—training‑time safeguards, real‑time misuse defenses, prompt‑injection resilience, and chain‑of‑thought monitoring—and pushes alignment further. Shortcut exploitation success drops to 0%, Terminal‑Bench 4.0 cheating attempts fall to 0.4%, while overall accuracy rises. Internal tests show unsupported claims cut by roughly half, with the model inserting methodological caveats or refusing out‑of‑scope actions. When faced with ambiguity, Carbon retrieves relevant sources and surfaces discrepancies before acting, enhancing collaboration. Over‑reach decreases by 26% and under‑service by 29%, yielding more reliable multi‑step plan execution. On the Gray Swan Indirect Prompt Injection benchmark, Carbon achieves a 0% attack success rate.
We aim to deliver meaningful leaps in software engineering and cybersecurity defense without compromising safety or reliability, and look forward to expanding access soon.
Review: Gemini 4 Carbon marks a notable advance in secure coding and long‑term reasoning for large models, coupling higher alignment and transparency with robust safety layers, making it a compelling candidate for real‑world cyber‑defense deployments.