Google announced Gemini 4 Carbon as its newest frontier model, emphasizing marked improvements in software development, autonomous task handling, and defensive security. In the release, the company said the system achieved the highest ranking on the Artificial Analysis Intelligence Index version 5.0, surpassing competitors such as Claude Opus 5.5 and GPT-6 Astra. The claim highlights a strategic push to advance AI capabilities beyond existing large-language offerings.
Independent testing shows Gemini 4 Carbon beating leading models from OpenAI and Anthropic on several evaluation suites. The system currently holds the top position on SEC-bench Pro, a benchmark that gauges an AI’s ability to locate authentic software bugs, as well as on ExploitGym, which measures how effectively a model can transform a discovered vulnerability into a functional exploit. These results suggest a notable edge in practical security analysis.
Access to the model is presently restricted to a handful of cybersecurity defense firms, with Google planning a broader rollout through its Fairwind Program. Pricing mirrors that of the earlier Argon release, set at $2 for each million input tokens and $10 for each million output tokens. This structure aims to lower entry barriers for enterprise adopters.
Google describes Carbon as engineered for extended, multi-step assignments, noting internal experiments where the model operated continuously for days, navigating dead-ends to resolve intricate engineering problems. The company also claims the system is more resistant to shortcut tactics and prompt-injection attacks, and that it collaborates more effectively with human operators. These attributes are intended to improve reliability when the model is deployed in high-stakes security environments.
The release also emphasizes greater transparency, with Google stating that Gemini 4 Carbon reduces unsupported assertions by a factor of two compared with its predecessor Argon. By cutting false or speculative outputs, the model aims to present more honest responses while maintaining performance. This focus aligns with broader industry calls for accountable AI behavior in critical domains.