Google's Gemini 4 Carbon takes top spot on Artificial Analysis Intelligence Index v5

Gemini 4 Carbon leads the Intelligence Index

Google's Gemini 4 Carbon has claimed first place on the Artificial Analysis Intelligence Index v5, according to the independent benchmarking service, posting an overall score of 59 with high reasoning enabled. The same evaluation puts the model second only to Anthropic's Claude Opus 5.5 on coding agents, an area where Google has historically underperformed.

Top of the Intelligence Index

Artificial Analysis reports that Carbon takes the number-one score in four of the nine evaluations that feed the index: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Its overall lead rests on complex reasoning, knowledge and scientific tasks, with substantial gains in coding and knowledge work relative to earlier Gemini models — although the firm notes it does not lead in every one of those areas.

The coding gap with Anthropic narrows

Paired with Google's Antigravity CLI, Gemini 4 Carbon at high reasoning scored 59 on the Artificial Analysis Coding Agent Index, three points behind Claude Code running Claude Opus 5.5 at maximum reasoning, which scored 62. Carbon also ranked second on both Terminal-Bench 4.0, at 64%, and FrontierCode v1.1, at 51%, again behind Claude Opus 5.5. Artificial Analysis singles this out as notable given Google's past weakness in agentic coding.

Knowledge work improves but still trails

On AA-Briefcase, the service's agentic knowledge-work benchmark, Carbon at high reasoning reached 1609 Elo — a 126-point improvement over Gemini 4 Argon's 1483 — yet remains behind Grok 4.7 at 1638, Muse Wonder at 1659 and Kimi K3.1 Pro at 1678. The largest single gain came on presentations, up 293 Elo to 1598, with analytical quality up 120 points to 1671.

Lower cost per task, higher hallucination than Argon

Carbon tops the recently updated AA-Omniscience v1.1 with a score of 54, six points ahead of Claude Opus 5.5 and eight ahead of Argon, with accuracy up ten points over Argon to 58%. Its hallucination rate rose from Argon's 4% to 11%, though Artificial Analysis says that is still the second-lowest of any model, behind only Argon, compared with 50% for both Claude Opus 5.5 and GPT-6 Astra at max reasoning.

At standard pricing, Carbon costs $7.76 per Intelligence Index task, roughly 23% less than Claude Opus 5.5's $10.04, but about 1.2 times GPT-6 Astra's $6.25. It uses 123k output tokens per task, 29% fewer than Opus. Priced identically to Argon at $4 per million input tokens and $20 per million output tokens, Carbon scores nine points higher on the index at essentially the same cost per task as Argon's $7.68. Cached input receives a 95% discount, working out to $0.20 per million tokens at standard rates. The model offers a 1M-token context window and accepts text, image, video and speech input, with text output.

Limited rollout

As with the launch of Gemini 4 Argon, Carbon is being released to selected users and is not yet publicly available. Google is offering a 50% launch discount with no confirmed end date, and Artificial Analysis says all of its charts and figures use standard pricing for both Gemini 4 models.

Why it matters

An independent benchmark putting a Google model at the top of general intelligence — while it simultaneously closes a long-standing deficit in agentic coding against Anthropic — shows how quickly frontier-model leadership is cycling among the major labs. The economics matter as much as the scores: comparable overall intelligence at roughly a quarter less cost per task, and with fewer output tokens per task, adds pressure on rival API pricing. One caution is that these results come from a model most developers cannot yet access, and the rise in hallucination rate from 4% to 11% suggests the intelligence gains carry a trade-off. Until Carbon reaches general availability, buyers should treat the numbers as early-access performance rather than something they can evaluate themselves today.