Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5

What it claims

Artificial Analysis (28 October 2026) ranks Gemini 4 Carbon, run with high reasoning, first on its Intelligence Index v5 with a score of 59. It takes the #1 score on 4 of 9 Index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. The article states that Carbon does not lead in coding or agentic knowledge work, though it gains substantially over earlier Gemini models in both.

Evidence

Coding agents. With Antigravity CLI, Carbon (high) scores 59 on the Coding Agent Index. Claude Code with Claude Opus 5.5 (max) scores 62. Carbon ranks second on Terminal-Bench 4.0 (64%) and on FrontierCode v1.1 (51%). The page gives Opus 5.5 (max) a figure of 56%, apparently for FrontierCode. The Opus 5.5 Terminal-Bench number is not stated.

Agentic knowledge work. On AA-Briefcase, Carbon (high) scores 1609 Elo, 126 points above Gemini 4 Argon (high, 1483). It trails Grok 4.7 (xhigh, 1638), Muse Wonder (max, 1659) and Kimi K3.1 Pro (1678). Presentation rises 293 Elo to 1598, and Analytical Quality rises 120 points to 1671.

Knowledge and hallucination. On AA-Omniscience v1.1, Carbon scores 54. Claude Opus 5.5 (max) scores 48 and Gemini 4 Argon scores 46. Accuracy rises 10 points over Argon to 58%. The hallucination rate also rises, from 4% (Argon) to 11%. That is still the second lowest of any model tested. Claude Opus 5.5 (max) and GPT-6 Astra (max) each have a 50% hallucination rate.

Cost per task. These figures use standard pricing:

Carbon costs about 23% less per task than Claude Opus 5.5 (max) and about 1.2x as much as GPT-6 Astra (max). It uses 123k output tokens per task, against 173k for Opus 5.5 (max). Carbon and Argon cost about the same per task, but Carbon scores 9 points higher on the Index.

How to use it

• Pricing: $4 per 1M input tokens and $20 per 1M output tokens, the same standard pricing as Gemini 4 Argon.
• Launch discount: Google offers 50% off, with no confirmed end date.
• Cached input: Cached input tokens get a 95% discount, which is $0.20 per 1M tokens at standard pricing.
• Context window: 1M tokens.
• Modalities: Text, image, video and speech input. Text output only.
• Reasoning setting: All results use high reasoning. The coding result uses the Antigravity CLI harness.

Limits

Carbon is rolling out to selected users only and is not publicly available. The article does not say how to get access.

All figures use standard pricing, so actual cost during the discount period is lower.

The hallucination rate is higher than Argon's (11% against 4%), even though accuracy improved.

The coding comparison mixes harnesses and settings: Antigravity CLI at high reasoning against Claude Code at max. The article gives no Carbon results at other reasoning levels.

Index v5 is a new version, announced on 27 October 2026. The page links to that announcement, which says most of the Index now runs on private datasets. Scores therefore cannot be checked independently and are not directly comparable to earlier Index versions.