Gemini 4 Carbon tops Intelligence Index v5 at $7.76 a task

A buyer picking between frontier models rarely sees a cost figure that is not the vendor's own. Artificial Analysis measured Google's Gemini 4 Carbon on its new Intelligence Index v5 at $7.76 per task. That is 23 percent less than Claude Opus 5.5, which scores second. The article was posted on 28 October and sits at 328 points on Hacker News.

Artificial Analysis tracks models with a fixed set of evaluations and reports its own numbers. Intelligence Index v5 combines nine tests. On the latest reading, Gemini 4 Carbon scores 59 overall. Claude Opus 5.5 is next. Gemini 4 Carbon takes first on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.

The cost gap

Artificial Analysis reports Gemini 4 Carbon used 123,000 output tokens on the task set. Claude Opus 5.5 used 173,000, which is about 29 percent more. At list prices the two came out at $7.76 and $10.04 per task.

The pricing on Gemini 4 Carbon is $4 per million input tokens and $20 per million output tokens. Google is running a 50 percent launch discount at the time of the measurement, Artificial Analysis says. Cached inputs are listed at $0.20 per million. The model has a 1 million token context window.

Where it loses

Gemini 4 Carbon does not win every test. On AA's Coding Agent Index it scores 59, three points behind Claude Opus 5.5 at 62. On knowledge work, Muse Wonder and Kimi K3.1 Pro lead. Artificial Analysis gives one metric where Gemini 4 Carbon pulls clear: a hallucination rate of 11 percent, which it calls the second-lowest it has measured. Claude Opus 5.5 is at 50 percent on the same test.

Reading the number

The 23 percent gap is one data point from one evaluator on one task set. A team running agent code should still look at the three-point Coding Agent Index gap. Long autonomous runs lean on that score. The useful figure here is not the ranking. It is the token count: 123,000 against 173,000 output tokens for the same work. That difference carries through every bill a buyer pays, no matter which model wins the next release cycle.

Google announced Gemini 4 Carbon on 27 October with a cybersecurity focus. The Artificial Analysis article is the first independent read on its cost and output.

Source

Artificial Analysis article on Gemini 4 Carbon. Hacker News discussion here.