Google’s Gemini 4 Carbon scored 59 on Artificial Analysis’ Intelligence Index v5, placing it first among the models tested. Artificial Analysis says Carbon led four of the index’s nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. The company says its strongest results were in complex reasoning, knowledge and scientific tasks, though Carbon did not lead every area.
The result does not mean Carbon is widely available. Google is rolling it out to selected users, and it is not yet publicly available. Its standard price is $4 per million input tokens and $20 per million output tokens, matching Gemini 4 Argon. Google is offering a 50% launch discount, but has not confirmed when it will end.
How did it do on coding and office work?
In Artificial Analysis’ Coding Agent Index, Gemini 4 Carbon with high reasoning scored 59 when run through Antigravity CLI. That put it three points behind Claude Code with Claude Opus 5.5 in max mode, which scored 62. Carbon ranked second on Terminal-Bench 4.0 with 64% and FrontierCode v1.1 with 51%; Artificial Analysis says Claude Opus 5.5 was ahead on both.
Carbon also improved on Gemini 4 Argon in AA-Briefcase, Artificial Analysis’ test of agentic knowledge work. Carbon with high reasoning scored 1,609 Elo, 126 points above Argon with high reasoning. It still trailed Muse Wonder in max mode (1,659), Grok 4.7 in xhigh mode (1,638) and Kimi K3.1 Pro (1,678). Its biggest reported gain was in presentations, up 293 Elo to 1,598.
What do the scores and prices show?
Artificial Analysis says Carbon costs $7.76 per Intelligence Index task at standard pricing, about 23% less than Claude Opus 5.5 in max mode at $10.04, but more than GPT-6 Astra in max mode at $6.25. The benchmark’s task-cost estimate also puts Carbon close to Argon, at $7.68, even though Carbon’s index score is nine points higher. Carbon has a 1-million-token context window and accepts text, images, video and speech, with text output. These are Artificial Analysis’ benchmark results; the company’s evaluation does not establish how the model will perform for every user or task.
Why it matters
For people who can access Carbon, the reported scores suggest it may be a strong option for reasoning and scientific tasks, and its standard per-task cost in this evaluation was lower than Claude Opus 5.5’s. But most people cannot try it yet: Google has limited the rollout to selected users, and the end date for the discount is unknown.
