Google has announced its latest frontier model, Gemini 4 Carbon, claiming top benchmark scores in key areas such as coding and cybersecurity. The new release follows the launch of Gemini 4 Argon just two weeks prior, with the company highlighting the rapid pace of its development in a blog post.
Benchmark Performance and Capabilities
According to Google’s benchmark data, Gemini 4 Carbon achieved the highest score on the AA Index v5.0 general AI benchmark with 59, placing it slightly ahead of Claude Opus 5.5 and GPT-6 Astra. The model also demonstrated strong performance across multiple categories, including agentic coding, computer use, knowledge work, and science and math. Google drew particular attention to Carbon’s performance in cybersecurity, where it outperformed competing models by significant margins.
Google stated that the model excels at identifying, prioritizing, and patching vulnerabilities at their root cause. Additionally, the company noted that Carbon enhances safety by minimizing shortcuts and loophole exploitation, improving honesty and transparency, resisting prompt injection, and generating secure code twice as frequently as the Argon model.
Release Strategy and Rollout
Following industry precedent for frontier models, Google is initially restricting Gemini 4 Carbon access to select cybersecurity defense organizations for pre-release testing. Broad access will later be granted to trusted partners via the Fairwind Program prior to a general public release. This phased approach mirrors the rollout of Gemini 4 Argon, which is now accessible to the public through the Gemini app and AI search mode.
