Google has introduced its most powerful model, Gemini 4 Carbon

Google has introduced a new language model, Gemini 4 Carbon, which scored 59 points in the Artificial Analysis ranking, surpassing the previous leader. The main advantage of the new model is cybersecurity: it is capable of finding vulnerabilities in code, verifying their exploitability, and independently patching the discovered holes.

Due to its automated hacking capabilities, the model is not yet available to the general public. Initial access has been granted to specialized professionals, government agencies, and major platforms. In exploit tests, Carbon successfully completed 54.8% of tasks, outperforming its competitors. In the GraphWalks test with a long context of over a million tokens, the model achieved a score of 96.2%. In the external business test Vending-Bench 2, the AI was able to complete a simulated year with a balance of $16,484 in the vending machine. However, in tests involving computer operation and engineering tasks, the new model lagged behind counterparts from other developers.

According to the developers' internal assessment, the hallucination rate of the new model has been halved compared to its predecessor, although independent security verification remains impossible until the official release for all users. The high frequency of new version releases demonstrates the company's aggressive strategy in the field of artificial intelligence.