Google Unveils Gemini 4 Carbon, Its Most Powerful AI Model Yet
Translated & summarized from Now 14 by baba
Google DeepMind has launched Gemini 4 Carbon, its most powerful AI model, capable of continuous operation for days and excelling in complex coding and cybersecurity. The model ranks first globally on the Artificial Analysis AI Index and has demonstrated significant improvements in detecting and patching vulnerabilities, partly through collaboration with Israeli firm Wiz. Google is implementing a phased rollout, starting with cybersecurity professionals and government testing, before wider public release.
The story in 5 lines · by baba
- Google's Gemini 4 Carbon, its most powerful AI model, can operate continuously for days on complex tasks.
- The new AI model ranks first globally on the Artificial Analysis AI Index with a score of 59.
- Gemini 4 Carbon demonstrates leading cybersecurity capabilities, scoring high on vulnerability detection and patching.
- The model's advanced features were validated through collaboration with Israeli cybersecurity firm Wiz.
- Google is gradually releasing Gemini 4 Carbon, starting with cybersecurity experts and government testers.
Google DeepMind has unveiled its latest flagship AI model, Gemini 4 Carbon, which it claims is the world's most powerful. Announced by CEO Sundar Pichai via X (formerly Twitter), the model excels in complex coding, continuous task resolution, and cybersecurity defense. This release follows the launch of Gemini 4 Argon just two weeks prior, highlighting Google's rapid pace in AI development.
Gemini 4 Carbon is designed to function as an autonomous agent capable of running for days, unlike typical chatbots that lose context. Internal tests showed it analyzing vast code systems, overcoming dead ends, and identifying deep-seated issues in operating systems and infrastructure. Google Chrome teams reported a tenfold increase in efficiency for generating security patches and software fixes compared to competing models.
The model achieved a score of 59 on the Artificial Analysis v5.0 AI Index, ranking first globally. In cybersecurity evaluations, it scored 88.2% on SEC-bench Pro for vulnerability detection and 71% on CWE-bench for code patching. A significant portion of these evaluations involved collaboration with the Israeli cybersecurity firm Wiz, whose researchers confirmed Carbon's leading ability to uncover real-world vulnerabilities.
Google also highlighted breakthroughs in safety mechanisms, with Carbon showing 0.0% success rate in tests designed to bypass its instructions. The model's rate of unsubstantiated claims was halved compared to its predecessor, and it avoided attempts to cut corners in code tests. Commercially, Google is maintaining introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached tokens, before a planned increase to $4 and $20 respectively.
Gemini 4 Carbon is being rolled out gradually and responsibly. Initially, it is accessible to a select group of cybersecurity professionals through Google's Fairwind Program and is undergoing voluntary testing with the U.S. government. Wider access for API developers, business clients, and the general public through Gemini applications is planned, though a specific date has not yet been set.