MOUNTAIN VIEW, November 7, 2026 — Google’s Gemini 4 Carbon debuted on October 28 and immediately claimed the top spot on the Artificial Analysis Intelligence Index V5 with a score of 59, edging out Anthropic’s Claude Opus 5.5 by a single point. The model also posted the highest score to date on the new AA-Omniscience v1.1 hallucination benchmark, though that figure — 54 on a -100 to +100 scale — underscores how far frontier models remain from consistent reliability.
By the Numbers
Carbon’s 59 on the Intelligence Index V5 places it two points ahead of OpenAI’s GPT-6 Astra and nine points clear of Meta’s Muse Wonder Full. On the AA-Omniscience test, which tracks knowledgeable recall and hallucination rates, Carbon’s 54 beats the field by significant margins, according to Artificial Analysis data cited by Tom’s Hardware. The bottom line: even the best model still posts a net positive score barely above the midpoint of a scale where zero represents chance-level performance.
Investors will note the cost structure. Carbon runs at $7.76 per task on the Intelligence Index test, cheaper than Claude Opus 5.5 and Sonnet 5.5, and roughly half the $12.99 per task of Claude Fable 5.1. OpenAI’s GPT-6.1 Sol, however, undercuts Carbon by a factor of six, pricing at roughly one-sixth the cost per task. At the lower end of the Pareto frontier, Chinese open-weight models GLM 5.3-Flash and DeepSeek V4 Pro dominate the price-to-performance curve, with DeepSeek V4.1 Flash 1017 coming in 13 times cheaper than Carbon while scoring 32 on the Intelligence Index.
Cyber Defenders First
Google is restricting initial access to cyber defenders through its Fairwind program, with wider availability slated for later. The staggered rollout aligns with the company’s participation in the U.S. government’s voluntary pre-release testing process for frontier models. Google Senior Vice President Koray Kavukcuoglu called Carbon “our strongest model with state-of-the-art cybersecurity and long horizon capabilities” in a blog post published by Quartz.
The company reports that Carbon thwarted all attempts on the Gray Swan Indirect Prompt Injection benchmark and cuts unsupported claims at twice the rate of Gemini 4 Argon, which debuted two weeks earlier. Google also claims a 26% reduction in model overreach, with Carbon presenting sources and suggested actions rather than pursuing its own intentions when uncertainties arise.
Agentic Endurance and Blind Spots
Internal Google trials tracked Carbon persisting for days on complex, multi-step engineering problems, debugging subtle faults across large codebases and low-level systems. The model recovered from failed approaches without handing control back to a human operator, a capability Google frames as a shift from steering each step to reviewing outcomes.
On WebDev Arena, Carbon climbed 89 Elo points to 1,768, landing in fourth place behind three unnamed competitors. It hit 88.2% on SEC-bench Pro for vulnerability detection, a mark Google describes as state-of-the-art. However, Artificial Analysis data shows Carbon trailing Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on knowledge work tasks, and falling short of Claude Code paired with Claude Opus 5.5 on programming benchmarks.
The Figure That Matters
Carbon debuted two weeks after Gemini 4 Argon, a release cadence that forces enterprise teams to evaluate whether to retool for each launch or wait for claims to settle. Google’s blog post, authored by Kavukcuoglu, blends external leaderboard results with internal trials that lack published sample sizes, dates, or methods. The post names no rival model aside from Argon and offers no baseline Intelligence Index V5 score for earlier Gemini models, limiting the ability to measure progress over time.
Sources
- 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
- 10 things to know about Google's Gemini 4 Carbon frontier AI model
Photo: BoliviaInteligente / Unsplash (Unsplash License)
