Google Releases Gemini 4 Carbon Large Model, with a Focus on Enhancing Cybersecurity Capabilities

On October 29, Google recently officially launched its new-generation frontier AI large model, Gemini 4 Carbon. The model focuses on two core directions: cybersecurity and long-horizon complex task processing, and has achieved impressive results in multiple evaluations.

According to the introduction, in the Artificial Analysis Intelligence Index v5.0 intelligence index evaluation, Gemini 4 Carbon scored a composite 59 points, ranking first on the global list. The model is specifically optimized for highly difficult, multi-day continuous agent workflows, and can autonomously eliminate invalid paths and complete multi-step complex engineering work such as large codebases, root-cause localization of low-level system failures, and vulnerability patch writing. On the WebDev Arena front-end full-stack development evaluation leaderboard, the model improved by 89 Elo points over the previous generation product, scoring 1769 points and ranking fourth on the leaderboard.

Cybersecurity is the focus of this model iteration and upgrade. The model has reached a relatively high industry level across the entire security defense chain of vulnerability discovery, result verification, and vulnerability remediation. Internal test results show that, compared with the previous generation Gemini 4 Argon, Gemini 4 Carbon's success rate in outputting secure code has more than doubled. Comparative testing conducted by the Google Chrome team shows that the model's ability to produce high-quality vulnerability patches far exceeds that of frontier large models developed by other organizations.

At the level of safety alignment capability, the model inherits and optimizes the original multi-layer safety system. Evaluation data shows that its Terminal-Bench 4.0 test cheating success rate reached 0.0%, and the proportion of cheating attempts was compressed to 0.4%; in the internal programming environment, the number of high-severity misalignment alerts dropped by nearly 80%. The model's unfounded fabricated output problem has been reduced by about half compared with the previous generation. When evidence is insufficient, it can proactively explain the boundaries of its own capabilities, refuse non-compliant operations, and list the relevant missing information. When encountering vague user requirements, the model will first consult reference materials and alert the user to information discrepancies, rather than blindly inferring and executing the task. Statistics show that the model's over-execution rate dropped by 26%, and its under-service rate dropped by 29%. The reliability of complete execution of multi-step tasks has been enhanced, and it can abide by the established scope of the project. At the same time, this version can fully resist the indirect prompt injection attacks covered by the evaluation, with an attack success rate of 0.0%.

In terms of pricing, Gemini 4 Carbon's initial pricing is consistent with the previously released Gemini 4 Argon, and cached input Tokens can enjoy a 95% discount on the input price. During the promotional period, input is $2 per million Tokens and output is $10 per million Tokens; after the promotional period ends, the regular input price is $4 per million Tokens and the output price is $20 per million Tokens.

According to the release arrangement, the model will cooperate with the U.S. government's voluntary pre-release model testing process, and this week priority trial access will be opened to a small number of cybersecurity protection practitioners. Subsequently, Google will rely on the Fairwind program to gradually expand access for trusted partners, and after fully collecting feedback and polishing and improving the security protection mechanisms, it will open use to developers, enterprises, and ordinary users at large.