Google reveals new Gemini 4 Carbon AI model, narrowly beating Claude Opus 5.5 - frontier AI model is also the least likely to hallucinate, benchmarks claim

Google has released Gemini 4 Carbon, a frontier AI model that tops several third-party benchmark rankings while posting the lowest hallucination score measured to date. According to results from Artificial Analysis, the model edges out Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on generalized intelligence and leads the field on the firm's knowledge and hallucination index. The release marks a comeback for Google, which had a year of irregular model launches, and arrives as businesses and government agencies weigh reliability alongside raw capability.

Gemini 4 Carbon (high) scored 59 on the Artificial Analysis Intelligence Index V5, one point ahead of Claude Opus 5.5 (max with fallback) and two ahead of GPT-6 Astra (max). It finished nine points ahead of Meta's Muse Wonder Full (Max) and 10 ahead of Claude Fable 5.1 (max with fallback). On the AA-Omniscience v1.1 index, which runs from -100 to +100, Carbon scored 54, six points above Claude Opus 5.5 (max) at 48. Even that top score suggests frontier models in 2026 remain far from consistently reliable at recalling information accurately.

Price is a central part of the pitch. Artificial Analysis measured Gemini 4 Carbon at $7.76 per task on the Intelligence Index, slightly below Gemini 3.8 Flash (high) and about 23% less than Claude Opus 5.5 (max) at $10.04. That is several dollars cheaper than Claude Opus 5.5 or Sonnet 5.5, both max with fallback, and not far off half the $12.99 per task recorded for Claude Fable 5.1. Carbon remains pricier than OpenAI's GPT-6 Astra (max) at $6.25 and roughly 13 times the cost of the open-weight DeepSeek V4.1 Flash 1017, which scored 32 on the index.

The model is not a leader in every category. Artificial Analysis found Gemini 4 Carbon trails Claude Code with Claude Opus 5.5 at programming and falls behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work. Arena.ai places Carbon atop its Pareto Frontier chart at the higher end of the price spectrum, while lower price points are held by Chinese open-weight models including GLM 5.3-Flash and DeepSeek V4 Pro.

Google said it is rolling out Carbon gradually and is participating in the U.S. government's voluntary pre-release testing process. Access is opening first to a subset of cyber defenders, with wider availability through trusted partners under its Fairwind Program, while Google iterates on guardrails intended to keep the model from being used for exploits and hacking. Google also said Carbon thwarted all attempts under the Gray Swan Indirect Prompt Injection benchmark to make it act against instructions.

On reliability, Google claims Carbon cuts unsupported claims in its responses by twice as much as Gemini 4 Argon, introduced two weeks earlier, and that it declines to act rather than fabricating answers or hedging when evidence is missing. Google says the model more readily presents sources and suggested actions and cut model overreach by 26%. Artificial Analysis recorded 58% accuracy for Carbon versus 48% for Argon, with a higher hallucination rate of 11% against 4%. The emphasis on cost comes as Gartner research shows just 3% of companies are achieving cost and productivity improvements from AI use.

More AI news from TechManNews.