Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
B · Useful, limited A short launch analysis by independent evaluator Artificial Analysis with its own measured scores, costs and hallucination rates for Google's Gemini 4 Carbon; first-hand numbers, but brief and with methodology on other pages.
Summary based on the full text·Original: about 4 minutes to read
In short
An Artificial Analysis article (28 October 2026) reporting that Google's Gemini 4 Carbon, at high reasoning, now tops its Intelligence Index v5 with a score of 59. It leads four of the nine index evaluations, is three points behind Claude Opus 5.5 on coding agents, and is not yet leading on agentic knowledge work. It is about 23% cheaper per task than Opus 5.5 at standard pricing. The model is rolled out to selected users only.
Key numbers
- 59Gemini 4 Carbon (high) score on the Intelligence Index v5, the top score
- $7.76 vs $10.04cost per Intelligence Index task at standard pricing, Gemini 4 Carbon against Claude Opus 5.5 (max)
- 59 vs 62Coding Agent Index score with Antigravity CLI against Claude Code with Opus 5.5 (max)
- 11%hallucination rate on AA-Omniscience v1.1, up from 4% for Gemini 4 Argon, against 50% for Opus 5.5 and GPT-6 Astra
- 123k vs 173koutput tokens per task, Gemini 4 Carbon against Claude Opus 5.5 (max)
As reported by the source.
English·Web page·opens artificialanalysis.ai
What it reports
- First place on 4 of 9 index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.
- Coding Agent Index 59 with Antigravity CLI, against 62 for Claude Code with Opus 5.5 (max); second on Terminal-Bench 4.0 at 64%.
- AA-Briefcase 1609 Elo, 126 points above Gemini 4 Argon but behind Grok 4.7 (1638), Muse Wonder (1659) and Kimi K3.1 Pro (1678).
- AA-Omniscience 54, six points ahead of Opus 5.5; accuracy 58%, hallucination rate 11%, second lowest of any model.
- Cost per task $7.76, about 1.2 times GPT-6 Astra (max, $6.25) and in line with Gemini 4 Argon ($7.68) while scoring 9 points higher.
- Standard price $4/$20 per million input/output tokens, with a 50% launch discount of no confirmed end date; cached input at a 95% discount. 1M-token context.
How they got it
Artificial Analysis runs its own benchmark suite on each model and computes cost per task from token use and standard list prices; all figures here use standard, not discounted, pricing.
Limits
- A short summary; per-benchmark methodology is on separate pages.
- The model is not publicly available, so results cannot yet be reproduced by others.
- Scores are for the high reasoning setting only.
Why it matters for companies
The top of the leaderboard now changes hands every few weeks, and the gaps between the leaders are a few index points. The more useful differences are cost per task and hallucination rate: on this evaluator's knowledge test, Gemini 4 Carbon gave wrong answers far less often than the other frontier models. Firms choosing models for factual work should weigh that rate and test on their own data, rather than follow the overall rank.
Summary based on the full text. Read the original for the full argument and data.Read the original on artificialanalysis.ai