Overview
- Google announced Gemini 4 Carbon Wednesday, Oct. 28, calling it its strongest Gemini 4-era model with improved long‑horizon agent workflows for multi‑step engineering and debugging tasks.
- The company published benchmark claims that place Carbon at the top of the Artificial Analysis Index (score 59) and report cybersecurity results of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench.
- Carbon will roll out in a staged program starting with a subset of cybersecurity defenders and trusted partners through Google’s Fairwind Program while the company completes guardrail iteration and voluntary U.S. pre‑release testing.
- Google says internal safety tests show large drops in unsupported claims, misalignment alerts and prompt‑injection success (including a reported 0.0% on Gray Swan’s IPI test), but most performance and safety figures come from company‑run evaluations and wider independent verification is pending.
- The release follows Argon by roughly two weeks, accelerates Google’s frontier cadence, and could reshape defender workflows by automating vulnerability discovery, exploit validation and patching while raising questions about dual‑use risks and how quickly broader access will expand.