Google’s Gemini 4 Carbon has taken the top spot on Artificial Analysis’ Intelligence Index V5, scoring 59 points and finishing one point ahead of Claude Opus 5.5. The model ranked first in four of the benchmark suite’s nine evaluations, according to the analysis. Google released it last week and is initially making it available to selected cyber defenders through its Fairwind program, with broader access planned later.

The model also led Artificial Analysis’ AA-Omniscience v1.1 evaluation, which measures knowledge recall and hallucinations. It scored 54 on a scale from -100 to +100. While that was the strongest result in the comparison, the score indicates that even highly ranked models can still struggle to recall information reliably.

Gemini 4 Carbon’s Intelligence Index score puts it two points ahead of OpenAI’s GPT-6 Astra and nine points ahead of Meta’s Muse Wonder Full. Artificial Analysis also reports a cost of $7.76 per task on the benchmark, slightly below Gemini 3.8 Flash (high) and several dollars less than Claude Opus 5.5 or Sonnet 5.5 in their tested configurations. It remains more expensive than some OpenAI models and substantially more expensive than the open-weight DeepSeek V4.1 Flash 1017, which scored 32 on the index.

The results do not put Carbon ahead in every area. Artificial Analysis says it trails Claude Code with Claude Opus 5.5 in programming, and falls behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work. Arena.ai’s Pareto Frontier chart places Carbon at the top among more expensive models, while lower-cost options include Chinese open-weight models such as GLM 5.3-Flash and DeepSeek V4 Pro.

Google says it is taking a gradual approach to the rollout and is participating in the U.S. government’s voluntary process for pre-release model testing. The company plans to provide access to a subset of cyber defenders first, then expand it through trusted partners in the Fairwind program. Google says it wants to refine the model’s safeguards before making it available to developers, with a focus on limiting misuse in cybersecurity.

The company also says Carbon resisted all attempts in the Gray Swan Indirect Prompt Injection benchmark to make it act improperly. Google claims the model produces half as many unsupported claims as Gemini 4 Argon, introduced two weeks earlier, and reduces model overreach by 26%. These are company claims; the available benchmark results do not mean the model is free from errors.

Sources: Tom's Hardware.