Google debuts flagship Gemini 4 Carbon model with cybersecurity upgrades

Google unveiled its latest flagship AI model, Gemini 4 Carbon, complete with enhanced cybersecurity measures. The announcement came just two weeks after the debut of its predecessor, Gemini 4 Argon. According to Google, Carbon performs better than its predecessor, achieving more than double the success rate in generating secure code.

Moreover, it outperforms GPT-6 Astra and Claude Opus 5.5 in various benchmarks, with a standout result on the Artificial Analysis Intelligence Index - a test combining over a half-dozen popular LLM evaluations. This index, released publicly today, features two additional benchmarks and removes some easy questions deemed too simple for advanced models.

Carbon excels at cybersecurity tasks, surpassing previous records on SEC-bench Pro, a test of an LLM's ability to identify vulnerabilities in complex software systems. It also patches the vulnerabilities it discovers more effectively than before. Koray Kavukcuoglu, Google's chief AI architect, noted that Carbon generated high-quality patches over 10 times better than another lab's large frontier model.

Google bolstered Argon's existing security features in Carbon and observed a nearly 80% reduction in misalignment alerts from internal teams using the model in coding environments. To prevent hallucinations - when LLMs fill knowledge gaps with inaccurate information - Google configured Carbon to avoid this behavior, reducing unsupported claims by about two-thirds compared to Argon.

Even during training, Carbon showed restraint, successfully completing only 0.4% of challenges in the Terminal-Bench 4.0 benchmark, with none of its attempts to cheat on the tasks. Google intends to release Carbon through its Fairwind Program later in the week, allowing select organizations to test it for cybersecurity improvements.

The company plans to refine Carbon's security protocols based on early user feedback before a broader release.