Google AI has unveiled Gemini 4 Carbon, its most advanced model yet, designed to push the boundaries of cybersecurity and long-horizon capabilities. Carbon ranks #1 on the updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59, showcasing significant improvements in coding, agentic tasks, and cybersecurity defense. It leads key benchmarks, including SEC-bench Pro (88.2%) for vulnerability detection, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. Internally, Carbon demonstrates a strong uplift in generating secure code and reducing misalignment and fabrication rates.
Building on the success of Gemini 4 Argon, launched just two weeks ago, Carbon represents the rapid pace of innovation at Google AI. The model is optimized for sustained agentic workflows, capable of tackling complex, multi-step engineering tasks over extended periods. It has shown remarkable performance in live coding environments, with internal teams reporting a more than 2x success rate in generating secure code compared to Argon. The Chrome team noted a 10x improvement in generating high-quality patches, while Wiz observed leading performance in web penetration testing.
Gemini 4 Carbon also introduces state-of-the-art cybersecurity capabilities, excelling across the defense lifecycle from vulnerability detection to remediation. The model has been trained to respond appropriately to complex tasks, reducing shortcuts and loophole exploitation. It cuts unsupported claims by approximately 2x over Argon and improves collaboration with users by surfacing discrepancies before acting. Additionally, Carbon achieves a 0.0% attack success rate on the Gray Swan’s Indirect Prompt Injection Benchmark, resisting every evaluated attack attempt.
Google AI is taking a deliberate approach to rolling out Carbon, starting with a subset of cyber defenders this week and gradually expanding access through the Fairwind Program. The model will be available at the same introductory price as Argon, with input tokens priced at $2 per million and output tokens at $10 per million. Cached input tokens are offered at a 95% discount. The company aims to bring Carbon to developers, businesses, and consumers as soon as possible, following further feedback and iterations on guardrails.