Google's Gemini 4 Carbon has moved to first place on version five of the Artificial Analysis Intelligence Index, according to results published by the independent model-testing company. The model recorded an overall score of 59 and led four of the index's nine component evaluations, covering difficult reasoning, science, broad knowledge and automation tasks.
The result does not make Carbon the leader in every category. Artificial Analysis placed it second on its Coding Agent Index, three points behind Claude Opus 5.5 when the models were used through their respective coding tools. Carbon also ranked second on Terminal-Bench 4.0 and FrontierCode v1.1. Those results nevertheless represented a notable improvement for Google's model family in agentic software work.
Artificial Analysis also reported gains in knowledge-work testing. Carbon reached 1,609 Elo on AA-Briefcase, 126 points above Gemini 4 Argon, although it remained behind several competing systems. Its largest subcategory improvement was in presentation work. The evaluation provider said Carbon accepts text, image, video and speech inputs and produces text output.
Cost and token use formed another part of the comparison. At standard prices, Artificial Analysis calculated a cost of $7.76 per Intelligence Index task for Carbon, about 23 percent below Claude Opus 5.5 in the tested configuration but above GPT-6 Astra. Carbon used 123,000 output tokens per task in that test, compared with 173,000 for Opus. Such figures depend on the benchmark harness, reasoning settings and provider pricing, so they should be read as results for this evaluation rather than universal measures of operating cost.
On the updated AA-Omniscience evaluation, Carbon scored 54. Artificial Analysis measured 58 percent accuracy and an 11 percent hallucination rate. That hallucination figure was higher than Gemini 4 Argon's 4 percent but lower than the rates measured for most other systems included in the comparison.
Access remains limited. Artificial Analysis said Google was rolling Carbon out to selected users and that it was not yet generally available. Standard pricing was listed at $4 per million input tokens and $20 per million output tokens, matching Argon, with a launch discount of 50 percent and no announced end date. Cached input tokens receive a deeper discount.
The ranking offers a snapshot of performance under a defined set of tests, not a complete assessment of reliability or suitability for a particular deployment. Carbon's overall lead, paired with narrower second-place results in coding and continuing gaps in agentic knowledge work, shows why buyers and developers need to examine task-specific scores alongside a headline index position.
