Google Launches Gemini 4 Carbon AI Model with Top Intelligence Ranking
October 31, 2026 · 3 min read
Google has released Gemini 4 Carbon, a next generation frontier AI model designed for high performance in cybersecurity and long term autonomous tasks. The model was announced on October 28, arriving only two weeks after the release of the Gemini 4 Argon model.
Gemini 4 Carbon ranked first globally on the Intelligence Index 5.0 by Artificial Analysis with a score of 59. Google reported that the model shows significant performance improvements over its predecessor in coding and agentic tasks. In the WebDev Arena for web development, the model scored 1768 points, which is an increase of 89 points, placing it fourth.
The model demonstrates advanced cybersecurity capabilities, including the ability to detect vulnerabilities, verify exploitability, and generate fix codes. It scored 88.2 percent on SEC-bench Pro for vulnerability detection, 54.8 percent on ExploitGym for exploit verification, and 71.0 percent on CWE-bench for vulnerability correction. Internal evaluations showed that the success rate for generating secure code is more than twice that of Gemini 4 Argon, while tests by the Google Chrome development team indicated that the performance for generating high quality security patches is more than 10 times higher than other large frontier models from different AI research institutions. Wiz, a cloud security platform, also confirmed the model's high performance in web penetration testing evaluations.
Gemini 4 Carbon is optimized for long term complex problem solving. Google stated that internal tests showed the model could continue working for several days to analyze and fix complex errors in large codebases and low level systems. Unlike previous models focused on single questions or limited code, this model is designed to perform multi step engineering tasks and find solutions when problems occur during the process.
Regarding safety and reliability, the model was trained to prevent cheating in evaluation environments. In Terminal-Bench 4.0, the cheating success rate was 0.0 percent and the attempt rate was 0.4 percent. In actual internal coding environments, high severity alignment mismatch warnings decreased by about 80 percent, and the frequency of unfounded claims was reduced by about half compared to Gemini 4 Argon. Over-action decreased by 26 percent and under-action decreased by 29 percent. In Gray Swan indirect prompt injection evaluations, the attack success rate was 0.0 percent, compared to 0.7 percent for Gemini 4 Argon.
Due to the risk that cybersecurity capabilities could be misused, Google is using a phased release. Access is being provided this week to trusted cybersecurity defense experts, with future expansion to verified partners through the Fairwind Program. Google is also participating in voluntary pre model evaluation procedures by the United States government.
Initial API pricing is set at 2 dollars per 1 million input tokens and 10 dollars per 1 million output tokens, matching Gemini 4 Argon. Cached input tokens receive a 95 percent discount. After the initial period, prices will adjust to 4 dollars per 1 million input tokens and 20 dollars per 1 million output tokens.