Artificial Analysis: Gemini 4 Carbon tops its index at $7.76 per task

Artificial Analysis, an independent benchmarking firm, said on 28 October 2026 that Google's Gemini 4 Carbon is the new number one on its Intelligence Index v5, with a score of 59. It also published cost and hallucination figures that Google's own launch post leaves out, and some of them are less flattering than the headline rank.

Where Carbon leads, and where it does not

Artificial Analysis reports that Carbon holds the top score on 4 of the index's 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It says Carbon is not leading in coding or knowledge work.

• Coding agents: with Google's Antigravity CLI, Carbon scores 59 on the firm's Coding Agent Index, 3 points behind Claude Code running Claude Opus 5.5 at max effort (62).
• Knowledge work: on AA-Briefcase, Carbon scores 1609 Elo, up 126 from Gemini 4 Argon, but behind Grok 4.7 (1638), Muse Wonder (1659) and Kimi K3.1 Pro (1678).

What a task costs

At standard pricing of $4 per million input tokens and $20 per million output tokens, the firm puts Carbon at $7.76 per Intelligence Index task. That is about 23% below Claude Opus 5.5 at max effort ($10.04) and about 1.2 times GPT-6 Astra at max effort ($6.25). Carbon used 123,000 output tokens per task, 29% fewer than Opus 5.5's 173,000.

Google's launch price is a 50% discount that Artificial Analysis says has no confirmed end date; its charts use the standard price. Argon costs $7.68 per task on the same measure, so the 9-point gain over Argon comes at almost no extra cost per task.

The hallucination trade-off

On AA-Omniscience v1.1, Carbon scores 54, ahead of Opus 5.5 (48) and Argon (46). Its accuracy rose 10 points over Argon to 58%. Its hallucination rate rose from 4% to 11%. The firm says that is still the second lowest of any model it has tested, behind only Argon, against 50% for both Opus 5.5 and GPT-6 Astra.

For a team that would rather get an answer of "I don't know" than a confident wrong one, that increase from 4% to 11% is the number to check against your own documents.

What the report does not settle

• Carbon is not publicly available. Artificial Analysis says it is rolling out to selected users, so no one outside that group can repeat these runs.
• Per the firm, the majority of Index v5 now rests on private datasets, so outsiders cannot inspect the questions.
• Task cost depends on the task mix. A figure of $7.76 per index task is not a price for your workload.

Google's own announcement, covered earlier on this site, cites the same first-place rank, which makes Artificial Analysis the source for both claims. The costs, the Coding Agent Index and the hallucination figures here come only from Artificial Analysis.