Details
- Google announced Gemini 4 Carbon on October 28, describing it as its strongest model to date.
- The model targets autonomous coding, cybersecurity defense, agentic workflows, scientific tasks, full-stack web development, and multimodal understanding.
- Google reported scores of 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
- Carbon is designed for long-horizon work, allowing agents to handle complex, multi-step engineering tasks over extended periods.
- Internal testing showed more than twice the success rate of Gemini 4 Argon for generating secure code, according to Google.
- Access begins this week for a subset of cyber defenders through Google’s Fairwind Program, followed by a gradual expansion to trusted partners.
- Broader availability for developers, businesses, and consumers is planned, but Google did not provide a public release date.
Impact
Carbon intensifies competition among Google, OpenAI, and Anthropic around coding agents and cybersecurity automation. Its restricted rollout suggests Google is balancing demand for more autonomous systems with misuse and reliability risks, particularly because the model can identify vulnerabilities, test exploits, and generate patches. The cybersecurity-first distribution also gives Google a controlled way to validate safeguards before wider access, while reported gains over Argon could encourage organizations to shift from general-purpose assistants toward specialized engineering agents.