Google introduced Gemini 4 Carbon, describing the new Google DeepMind model as its strongest to date, with major improvements in cybersecurity, coding, reasoning, long-horizon agentic workflows, and model safety.
Gemini 4 Carbon builds on Gemini 4 Argon, which Google launched two weeks earlier, and is designed to tackle more difficult engineering problems that require sustained work over long periods.
Google said internal teams have observed Carbon running for days while working through complex multi-step engineering tasks, including identifying root causes and patching subtle bugs across large codebases and low-level systems.
Carbon currently ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.
The model also posted significant gains across several cybersecurity benchmarks. Gemini 4 Carbon achieved 88.2% on SEC-bench Pro for identifying real-world software vulnerabilities, 54.8% on ExploitGym for validating vulnerabilities through exploitation, and 71% on CWE-bench for generating working fixes for identified security issues.
Google said Carbon achieved more than twice the success rate of Gemini 4 Argon on internal secure-code generation benchmarks. The Chrome team also reported that Carbon performed more than 10 times better than another frontier model when generating high-quality software patches.
Outside cybersecurity, Google said Carbon improved across coding, agentic tasks, scientific reasoning, multimodal understanding, and full-stack web development. The model gained 89 Elo points on WebDev Arena, reaching fourth place with a score of 1,768.
Google is taking a phased deployment approach.
The company is participating in the U.S. government’s voluntary pre-release model testing process and initially making Carbon available to a subset of cybersecurity defenders. Access will gradually expand to additional trusted partners through Google’s Fairwind Program as the company gathers feedback and refines safeguards.
Google plans to make Carbon available to developers, businesses, and consumers through its broader products afterward.
Gemini 4 Carbon will launch at the same introductory pricing as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens will cost 95% less than the standard input-token price.
Google also highlighted several safety improvements. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0%, while attempted cheating fell to 0.4%. Internal testing showed high-severity misalignment alerts down nearly 80% and unsupported claims down about half compared with Argon.
The company also said Carbon reduced model overreach by 26% and under-service by 29%.
On Gray Swan’s Indirect Prompt Injection benchmark, Carbon recorded a 0% attack success rate, resisting every evaluated attack attempt.
Google said the model is intended to combine increasingly autonomous engineering capabilities with stronger safeguards as the company expands access beyond its initial group of cybersecurity users.
