Gemini 4 Carbon takes MathArena lead from GPT-6.1 Sol at 15 times the cost

Gemini 4 Carbon takes the overall lead

MathArena, an independent benchmark for mathematical problem solving, added results for two frontier models on November 13, and the combined leaderboard has a new number one: Google's Gemini 4 Carbon displaces OpenAI's GPT-6.1 Sol from the top spot. Jasper Dekoninck of the MathArena team noted on X on November 14 that this is the first time in a long while that a non-OpenAI model holds first place. Anthropic's Claude Fable 5.2 took second, which means no OpenAI model currently tops the ranking.

The update reflects newly measured results rather than new releases. Google opened Carbon to all paying developers on November 10, and Anthropic launched Claude Fable 5.2 a day earlier, on November 9. Both models were already available; what changed is that MathArena has now scored them.

The numbers

According to the overall table cited by The Decoder on November 14, Gemini 4 Carbon in its high reasoning setting scores 85.4 percent with a 3.0 margin of error. Claude Fable 5.2 in max mode follows at 82.8 percent with a 5.3 margin, and GPT-6.1 Sol in max mode sits at 80.6 percent with a 2.7 margin.

On the ArXivMath track, Carbon's overall result is 94.28 percent with a 3.64 margin of error, first among the eight systems evaluated. The MathArena model page also records a September 2026 ArXivMath run at 96.23 percent.

The format matters here. MathArena publishes accuracy alongside a margin of error, cost per full run and output-token volume for every competition, from monthly ArXivMath and BrokenArXiv editions to other problem formats, and lists sampling parameters for each model. Given how close the top two scores actually are — 85.4 versus 82.8 — the uncertainty ranges are as informative as the headline figures.

The cost of accuracy

The leaderboard also prices performance. With the reasoning settings used in testing, an expected full Carbon run costs about 15.70 dollars, versus roughly 40 dollars for Claude Fable 5.2 and about 1 dollar for GPT-6.1 Sol. As The Decoder frames it, that makes Carbon roughly 15 times more expensive per run than GPT-6.1 Sol, with Fable 5.2 around 40 times.

The MathArena listing for Carbon shows a November 10 release date, a maximum output window of 1,048,576 tokens, and current pricing of 2 dollars per million input tokens and 10 dollars per million output tokens.

Why it matters

The significance is less the score itself than the shift at the top: an independent benchmark refreshed without any vendor announcement, and the number one position now belongs to a model that is not from OpenAI — something that has not happened in a long while. For teams choosing between models, the math track is a reasonable proxy for the multi-step analytical work agents perform over spreadsheets, filings and research material.

The cost column is equally useful. The top result costs an order of magnitude more per run than the model sitting at 80.6 percent, so a gap of a few points — one that comes with attached uncertainty ranges — has to be weighed against a per-run budget. The practical approach is to keep a benchmark-informed shortlist but measure on your own tasks: an automation budget feels the cost of a completed run, not a leaderboard point.