Google DeepMind releases Gemini 4 Carbon with top cybersecurity scores

Google DeepMind announced the release of Gemini 4 Carbon, describing it as their next frontier model. The company characterizes the new system as its strongest model, featuring state-of-the-art cybersecurity and long horizon capabilities. Gemini 4 Carbon is built on the predecessor model Gemini 4 Argon, which was launched two weeks prior to this release.

The model ranks number one on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. In specific cybersecurity evaluations, Gemini 4 Carbon achieved a score of 88.2 percent on the SEC-bench Pro benchmark for finding vulnerabilities. It recorded a score of 54.8 percent on the ExploitGym benchmark for exploitation and 71.0 percent on the CWE-bench benchmark for patching. Additionally, the model ranks fourth on the WebDev Arena with an Elo rating of 1768, following a jump of 89 Elo points.

Performance metrics indicate that Gemini 4 Carbon generates 2.3 times as many reasoning tokens and 64 percent more output tokens than Argon. In legal research tasks on Vals AI, the model averages more than 50 minutes per task. Pricing for the service is set at $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at 95 percent off the standard input token price.

Rollout of Gemini 4 Carbon begins this week to a subset of cyber defenders. Access will gradually expand to additional trusted partners through the Fairwind Program. Google DeepMind states that it is actively engaged in the U.S. government's voluntary process for pre-release model testing.