Gemini 4 Carbon edges Claude Opus 5.5 on a key AI index

Google's newest model has taken first place on the Artificial Analysis ranking, and it costs less per task than some other top models, including Claude Opus 5.5.

One point ahead on the Intelligence Index

Artificial Analysis, an AI benchmarking outfit, puts Gemini 4 Carbon at the top of its Intelligence Index V5. The model scores 59 in its "high" setting. Claude Opus 5.5 is one point behind, and OpenAI's GPT-6 Astra is two points back. Meta's Muse Wonder Full trails by nine points.

Carbon takes the best score on four of the index's nine evaluations. It is not best at everything. Artificial Analysis says it falls short of Claude Code with Claude Opus 5.5 at programming. It also sits behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work.

Cost per task

On the Intelligence Index, Carbon costs slightly less per task than Gemini 3.8 Flash, which is priced at $7.76 per task. It is several dollars cheaper than Claude Opus 5.5 or Sonnet 5.5. Claude Fable 5.1 costs $12.99 per task.

Cheaper options exist. Carbon costs about six times more than GPT 6.1 Sol and about 13 times more than the open-weight DeepSeek V4.1 Flash 1017, which scores only 32 on the index. At lower price points, Arena.ai shows open-weight Chinese models such as GLM 5.3-Flash and DeepSeek V4 Pro taking over.

The hallucination score

Carbon also tops AA-Omniscience v 1.1, a test of hallucinations and recall of facts. Its score is 54 on a scale from -100 to +100. That is the best result, yet it suggests that even leading models are far from reliably recalling information.

Google says Carbon cuts unsupported claims in its answers twice as much as Gemini 4 Argon, which it introduced two weeks earlier. The company also says the model shows sources and suggested actions more readily when evidence is uncertain, which it claims reduces overreach by 26%. These figures come from Google, not from an outside test.

A slow, supervised rollout

Google is opening Carbon first to a subset of cyber defenders through its Fairwind program. Wider access will follow through trusted partners. Developers have to wait. Google wants to tune the guardrails first, so the model is harder to misuse for exploits and hacking.

The company says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing." It also says the model resisted every attempt in the Gray Swan Indirect Prompt Injection benchmark.

Our view

A one-point lead on a single index can disappear with the next release, so the hallucination result may matter more to buyers. Organizations weighing AI spending care about reliability and cost per task. Whether Carbon holds its position may depend on how quickly Google widens access beyond Fairwind and on how rivals answer.

Source: Tom's Hardware