Google deploys Gemini 4 Carbon to top security benchmarks

Google released Gemini 4 Carbon on Wednesday, positioning the model as a direct challenger to OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 in the high-stakes arena of artificial intelligence security. The company stated that Carbon currently holds the number one spot on the Artificial Analysis Intelligence Index v5.0, marking a clear advantage over its two primary competitors in frontier model performance.

Security remains the focal point of the release. Google reported that Carbon leads the SEC-bench Pro evaluation, which tracks a model’s ability to identify real-world software bugs. It also tops ExploitGym, a test that measures how effectively an AI can convert a discovered vulnerability into a functional exploit. These results suggest a measurable improvement in the model’s capacity to defend against cyber threats compared to previous versions.

The model is not yet available to the general public. Google is limiting initial access to a select group of cybersecurity defense companies. The company plans to expand availability through its Fairwind Program, which manages partnerships for its AI infrastructure. This phased rollout allows Google to test the model’s reliability in high-risk environments before a wider launch.

Carbon replaces Argon, a model Google released earlier this month. The new version is designed to handle longer, more complex engineering tasks. Google noted that internal teams observed the model running for days to work through dead ends. It also improves on resistance to prompt injection attacks and reduces unsupported claims by half compared to its predecessor, offering greater transparency in its outputs.

Pricing remains consistent with the introductory rates set for Argon. Developers will pay $2 for every million input tokens and $10 for every million output tokens. This cost structure mirrors the competitive pricing strategy Google adopted for its previous frontier models.