Google Announces Gemini 4 Carbon Focusing on Cybersecurity and Long Tasks
Frontier AI models · AI assistants · AI coding
Google has announced Gemini 4 Carbon, highlighting cybersecurity exploit mitigation, patching, and multi-day task handling, leading the updated Artificial Analysis index with a score of 59. Access is currently limited to select cybersecurity defenders, with promotional pricing at $2 per million input tokens and $10 per million output tokens. Benchmarks show it lags competitors in certain software engineering tasks and exhibits an increased hallucination rate metric.
Google announced its Gemini 4 Carbon model on October 28, focusing on cybersecurity capabilities and handling complex multi-step tasks running over multiple days. According to Google and reports citing the benchmarks, Carbon scored 59 points under high reasoning on the revised Artificial Analysis Intelligence Index v5.0, leading in 15 of 17 reported benchmarks. Google reported gains across vulnerability discovery, exploitation, and patching relative to Argon, alongside improvements in secure code generation. However, some security evaluations were conducted without guardrails. Carbon still trails rival models in specific software engineering and knowledge benchmarks, such as FrontierSWE 2 and AA-Briefcase. Additionally, its hallucination rate metric in the AA-Omniscience test increased from 4% in Argon to 11%. Carbon is currently restricted to selected security defenders and trusted partners, with broader release pending.
Key takeaways
- Google unveiled Gemini 4 Carbon, reaching a top score of 59 on the Artificial Analysis Intelligence Index v5.0 under high reasoning.
- Google reported improved security metrics, including 54.8% on ExploitGym, 71.0% on CWE-bench, and a 0.0% indirect prompt injection success rate.
- Carbon still trails competitors in select software benchmarks like FrontierSWE 2, and its hallucination rate metric rose from 4% to 11% on AA-Omniscience.
- Availability is currently restricted to select cybersecurity defenders, with promotional pricing set at $2 input and $10 output per million tokens.
Source：INSIDE Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.