Google has announced Gemini 4 Carbon, its latest frontier AI model, with a focus on cybersecurity and autonomous software engineering.
Introduced just two weeks after Gemini 4 Argon, Carbon ranks first on the Artificial Analysis Intelligence Index v5.0. Google reports significant improvements in the model’s ability to discover software vulnerabilities and produce fixes.
Carbon is also designed to handle engineering assignments lasting several days. According to Google, internal teams have used the model to resolve difficult software issues with less human intervention.
Despite reported improvements in reliability and security, Google is limiting initial access to selected cyber defenders and trusted partners while additional safety testing continues.
Why It Matters: Gemini 4 Carbon demonstrates how autonomous AI could take on a greater role in enterprise technology operations, potentially reducing the time and resources required to resolve software vulnerabilities. Its capabilities also bring greater attention to the reliability and oversight needed before organizations entrust critical systems to AI agents.
- Cybersecurity Capabilities Extend Into Remediation: Carbon achieved 88.2% on SEC-bench Pro for vulnerability discovery and 71.0% on CWE-bench for generating security patches. Google also reported that the model more than doubled Argon’s secure-code generation success rate in internal testing. The external evaluations were conducted without guardrails, and performance in production environments has yet to be established.
- Autonomous Coding Tasks Can Continue for Days: Google reports that Carbon can work through extended engineering assignments, including diagnosing difficult software bugs across large codebases. Internal teams observed the model continuing through unsuccessful attempts to complete assignments. Carbon also climbed to fourth place on WebDev Arena following an 89-point Elo improvement.
- Reliability Improvements Target Unwanted Model Behavior: Google’s internal evaluations found an almost 80% reduction in high-severity misalignment alerts compared with Argon. The company also reported fewer unsupported claims and improved resistance to prompt injection. Carbon successfully resisted every attack included in the evaluated Gray Swan benchmark, although results under controlled testing conditions may differ from real-world performance.
- Initial Access Remains Restricted During Safety Testing: Google is making Carbon available to selected cybersecurity defenders through its Fairwind Program while participating in voluntary U.S. government pre-release testing. The company plans to expand access after gathering feedback and refining safeguards, although no general availability date has been announced.
- Pricing Matches the Previous Model: Carbon will launch at the same introductory rates as Gemini 4 Argon, starting at $2 per million input tokens and $10 per million output tokens. Cached inputs will receive a 95% discount. Google has not yet provided detailed production cost estimates for enterprise deployments.
Go Deeper -> Gemini 4 Carbon: our next frontier model – Google
