Gemini 4 Carbon Tops the Artificial Analysis AI Model Ranking With a Score of 59

Artificial Analysis is an independent lab that runs language models through identical tests and publishes aggregate rankings. The main one is the Intelligence Index: nine evaluations that add up to a single score. The fifth version of the index came out on October 28, 2026, and Google's Gemini 4 Carbon took first place with a score of 59. One caveat up front: the model is currently available only to select users, and there is no broad release. Google is offering a 50% launch discount with no announced end date, but the lab uses standard prices in all its charts.

Where Carbon wins and where it doesn't

Of the index's nine evaluations, Carbon took first place in four:

• HLE-Diamond: the hardest set of expert-level questions on science and knowledge;
• Terminal-Bench-Science: scientific tasks the model solves directly in the command line;
• AA-Omniscience: a test of factual knowledge and of whether the model knows not to make things up;
• AutomationBench-AA: workflow automation, where the model has to connect several services.

In the other five evaluations, Carbon improved substantially but does not lead. These include coding and office work.

Coding agents: Google is finally close to the leader

A separate ranking, the Coding Agent Index, measures not the model itself but the pairing of a model and an agent program. Carbon ran through Antigravity CLI, Google's own command-line agent, and scored 59. The leader is Claude Code with Claude Opus 5.5 at maximum reasoning effort, with 62 points. The gap is three points.

Carbon also placed second in two individual tests:

• Terminal-Bench 4.0: terminal tasks, where Carbon scores 64%;
• FrontierCode v1.1: difficult programming assignments, where Carbon scores 51%.

In both, Claude Opus 5.5 is ahead. For Google this is a notable shift: the company has historically lagged in coding.

Office tasks: progress, but no lead

AA-Briefcase evaluates how an agent handles "office" work: analytics, documents, presentations. Results are expressed as an Elo rating, as in chess. Carbon scored 1609, which is 126 points higher than Gemini 4 Argon (1483). Presentations improved the most: +293 to 1598. Analysis quality gained 120 points and reached 1671.

That is not enough for first place. Three models are ahead:

• Kimi K3.1 Pro: 1678;
• Muse Wonder (max): 1659;
• Grok 4.7 (xhigh): 1638.

Cost per task and answer honesty

The lab calculates how much it costs to run one Index task. The cost depends on the token price and on how many tokens the model "talks out" while reasoning.

• Gemini 4 Carbon: $7.76 per task, 123K output tokens;
• Claude Opus 5.5 (max): $10.04 and 173K tokens. Carbon is about 23% cheaper and writes 29% fewer tokens;
• GPT-6 Astra (max): $6.25. It is cheaper than Carbon, which costs about 1.2 times as much;
• Gemini 4 Argon: $7.68 at the same price of $4/$20 per million input/output tokens. Carbon costs almost the same but scores 9 points higher.

On the updated AA-Omniscience v1.1 test, Carbon scored 54. Claude Opus 5.5 scored 48, and Argon 46. Answer accuracy rose by 10 points to 58%. The trade-off is the share of fabricated answers, meaning cases where the model answers confidently instead of admitting it doesn't know: it climbed from 4% to 11%. That is still the second-best result among the tested models; only Argon does better. For Claude Opus 5.5 and GPT-6 Astra, this figure is 50%.

Technical specs

• Context window: 1M tokens.
• Accepts text, images, video and speech as input, and outputs text only.
• Price: $4 per million input tokens and $20 per million output tokens. Cached input tokens are 95% cheaper, i.e. $0.20 per million.

What this means for teams

If you are building coding agents, put Carbon on your test list as soon as it becomes available: it is close to Opus 5.5 in quality and noticeably cheaper per task. For projects where facts matter and fabrication must stay low, the trend matters too: Carbon is more accurate, but it fabricates answers three times as often as Argon. For presentations and office automation, Kimi K3.1 Pro and Muse Wonder still look stronger. Also keep in mind the 50% discount: it is temporary, so it is better to calculate your budget at standard prices.