On November 14, 2026, the third-party math benchmark MathArena updated its leaderboard: Google's Gemini 4 Carbon took first place with 85.4% expected performance, pushing OpenAI's GPT-6.1 Sol (80.6%) down to third, with Anthropic's Claude Fable 5.2 (82.8%) in second. It is the first time in a long while that a non-OpenAI model has topped this board — but the price is steep: Carbon costs $15.70 per problem, about 15 times what Sol costs.
What this benchmark actually measures
MathArena is run by Jasper Dekoninck, a researcher at ETH Zurich. Its distinguishing feature is that it tests models only on newly released math competition and research problems, so models cannot win by memorization; each model solves every problem four times and the average is taken. Carbon's score was achieved at its "high" reasoning setting, Sol's at "max." The leaderboard added results for Carbon and Fable 5.2 on November 13: Carbon only became generally available on November 10, Fable 5.2 was released on November 9, while Sol had held the top spot since its September 29 release. One caveat: Carbon's 85.4% carries a confidence interval of plus or minus 3.0 points — and Fable 5.2, which scored higher, is also more expensive, at $40.20 per problem.
Is the expensive lead worth it?
For anyone using models for math, the most useful numbers on the board are in the cost column: Sol costs $1.07 per problem and Kimi K3.1 Pro $3.70, against $15.70 for Carbon and $40.20 for Fable. Dekoninck himself cautioned that, because of the cost, Carbon was only fully run on the ArXivLean evaluation, and some of its scores are statistical estimates rather than measurements. Google has issued no announcement about the top ranking; Sundar Pichai and Demis Hassabis merely reposted the benchmark operator's post without comment. In other words, the first place is real — but it was bought with a high reasoning setting and a high inference bill.
How to read this board when choosing a model
The board is more of a snapshot than a verdict: it proves Carbon has genuinely top-tier ability at pure mathematical reasoning, but whether the extra 4.8 points are worth 15 times the money on everyday tasks depends on how hard the problems are and how big the budget is. Google's fast model, Hydrogen, takes a different route, built around value for money, while Fable 5.2's entry bar and pricing are a different calculation altogether. Once Carbon has measured scores on more evaluations and its pricing settles, the ranking will be worth another look.
