Google launches Gemini 4 Carbon for cybersecurity and long-horizon tasks

a long exposure of colored lights in the dark — Adrien / Unsplash

Google has launched Gemini 4 Carbon, a frontier artificial intelligence model designed to lead in cybersecurity and long-horizon agentic workflows. The initial release is available to a select group of cyber defenders before expanding to developers and businesses. The model builds upon the previously released Gemini 4 Argon and currently holds the top position on the Artificial Analysis Intelligence Index version 5.0 with a score of 59.

The model is engineered for agentic workflows that require continuous operation over extended periods. Internal testing by Google revealed that Carbon can operate for several days to solve multi-step engineering problems. These tasks include identifying root causes and patching bugs hidden within large codebases and low-level systems. The company also reported improvements in coding, agentic tasks, scientific work, full-stack web development, and multimodal understanding. On the WebDev Arena, the model's score increased by 89 Elo points, placing it fourth with a total of 1,768 points.

In cybersecurity, Carbon achieves state-of-the-art performance across the entire defense cycle, from vulnerability detection to risk verification and remediation. External benchmark tests conducted without guardrails show specific results. For finding real vulnerabilities, the model achieved 88.2 percent on SEC-bench Pro, discovering complex flaws in production software. For risk verification through exploitation, it scored 54.8 percent on ExploitGym, helping teams prioritize which vulnerabilities to fix first. For root-cause patching, it achieved 71.0 percent on the CWE-bench by Collinear, converting detection results into functional code fixes.

Internal Google tests indicate that Carbon has a success rate for generating secure code that is more than double that of Argon. The Chrome team reported that Carbon produces high-quality patches more than ten times better than a large frontier model from another vendor. Wiz, a cloud security platform, found that Carbon leads in web penetration testing benchmarks.

Regarding model safety, Carbon uses the same multi-layer safety stack as Argon, including model training, real-time misuse prevention, prompt injection resistance, and chain-of-thought monitoring. Google reported alignment results for Argon in its Safety and Security report. For Carbon, the successful cheating rate on Terminal-Bench 4.0 dropped to 0.0 percent, and cheating attempts fell to 0.4 percent. Severe misalignment alerts in real coding environments decreased by nearly 80 percent. The model also claims unsupported data about half as often as Argon, reduces overstepping task boundaries by 26 percent, and reduces incomplete work by 29 percent. The successful attack rate on the Indirect Prompt Injection Benchmark by Gray Swan dropped from 0.7 percent in Argon to 0.0 percent.

Google is rolling out Carbon cautiously and has participated in the voluntary pre-release model testing process of the US government. Initially, the model is available to a group of cyber defenders. It will then expand to selected partners in the Fairwind Program. Google will collect feedback and adjust guardrails before opening access to developers, businesses, and general consumers through its products. During the introductory period, Carbon is priced at 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input tokens are 95 percent cheaper than standard input tokens. After the introductory period, prices will adjust to 4 dollars and 20 dollars per million tokens respectively.