Google's new AI opens to paying customers, but no longer tops the independent ranking

Fifty-nine points. That is the score of Gemini 4 Carbon, Google's new model, on the intelligence index from Artificial Analysis, an independent company that runs all the major AIs through the same tests. It ranks first out of 60 models evaluated. Its lead is slim, though: Anthropic's Claude Opus 5.5 scores 58 points and OpenAI's GPT-6.1 Astra 57, according to figures reported by LeMagIT on November 2.

Google presents Carbon, which appeared on October 28 according to Artificial Analysis, as its most powerful model. It is not open to everyone: according to LeMagIT, early access is limited to Google's partners, including the U.S. government, and a few customers of its cybersecurity program. No launch date in Europe has been announced. Its little brother, Gemini 4 Argon, had its technical sheet published on October 13, according to LeMagIT.

First in the ranking, but not across the board

A second third-party evaluator, Vals AI, also places Carbon at the top of its general ranking, with 70.45%, just ahead of Opus 5.5 (70.18%), according to LeMagIT. The details are more nuanced, however. Across the nine tests in the Artificial Analysis index, Carbon comes first in only four. In coding and knowledge work, it "does not rank first," Artificial Analysis summarizes, as quoted by LeMagIT. On agent-driven coding tasks, it remains behind Claude Opus 5.5 used with the Claude Code tool.

And what about cybersecurity, its advertised specialty? Google claims that Carbon beats its competitors on vulnerability-hunting and attack-validation tests. These results come from the vendor. LeMagIT notes that the model has not been submitted to Artificial Analysis's cyber index.

Two and a half minutes before the first word

The Artificial Analysis page for the model reveals other limitations. Carbon takes an average of 147.83 seconds to produce the first word of its answer, compared with under five seconds for the median model. It is a model that "thinks" at length before writing, and that has a cost: to complete the index, it produced 200 million units of text (tokens), nearly three times the median.

Its public price is $4 per million input tokens and $20 per million output tokens, twice the median for comparable models. Per task, it averages $7.76, less than Opus 5.5 ($10.02), but far more than OpenAI's GPT-6.1 Sol, which scores 55 points for $1.27, according to Artificial Analysis measurements cited by LeMagIT.

What this means for a European business

For now, nothing: Carbon remains out of reach. The ranking mainly shows that the race at the top now comes down to a point or two, a margin that shifts from one test to the next. For most uses, cost per task and response time matter more than first place. Sundar Pichai, Google's CEO, said that lighter versions of Carbon were in the works, without giving a date. Those are the ones to watch.

Update of November 14, 2026

Gemini 4 Carbon, Google's most powerful model, is no longer limited to a few partners. Since November 10, it has been available to paying developers and businesses through the Gemini API, Google AI Studio and Gemini for business, according to Google's announcement. It is also rolling out gradually to Google AI Ultra subscribers, in the Gemini app and in AI Mode in search. Its launch price is $2 per million input tokens and $10 per million output tokens until November 30, then $4 and $20. Google does not specify which countries are covered and has not announced any lighter version.

In the Artificial Analysis ranking, Carbon has lost first place. On the current version of the index (v5.0.1), it scores 60 points and ranks fourth out of 80 models, just ahead of Claude Opus 5.5 (59 points). It now trails Claude Fable 5.2, the model Anthropic released on November 9, which reaches 65 points at its highest setting, according to Artificial Analysis. The page dedicated to Carbon still shows no cybersecurity results.