Google LLC has launched Gemini 4 Carbon, a new large language model designed to excel at complex programming and cybersecurity tasks, according to Silicon Angle.
The release follows the debut of Gemini 4 Argon by only two weeks. Google reports that Carbon is more than twice as effective at generating secure code compared to its predecessor and outperforms competitors like GPT-6 Astra and Claude Opus 5.5. The model achieved a record score on the Artificial Analysis Intelligence Index, which was updated to include more rigorous testing criteria. Carbon demonstrates significant proficiency in identifying and patching software vulnerabilities, with the Chrome team noting it performed 10 times better than other frontier models in generating high-quality patches.
Furthermore, Google integrated enhanced guardrails to reduce hallucinations and misalignment, resulting in an 80% decrease in alerts during internal testing. The model also showed resilience against prompt injections, blocking all simulated attempts during evaluations. Google plans to provide early access to the model through its Fairwind Program later this week, allowing organizations to test its capabilities in live environments. Feedback from these participants will help refine the model before a broader release, marking a significant step in improving automated software security.
Source: Silicon Angle
