Gemini 4 Carbon takes the overall MathArena crown from GPT-6.1 Sol as the first non-OpenAI leader in a while

dev.to Machine Learning2026-11-15 01:550 views

📅 November 14, 2026 • MathArena

Independent math benchmark MathArena added results for two frontier models on November 13, and the overall picture shifted: Google's Gemini 4 Carbon now leads the combined leaderboard, displacing OpenAI's GPT-6.1 Sol from the top spot. Jasper Dekoninck of the MathArena team noted on X on November 13 that this is the first time in a long while a non-OpenAI model holds the number one position. Anthropic's Claude Fable 5.2 took second place, so no OpenAI model currently tops the ranking.

The numbers behind the shift: per the overall table cited by The Decoder on November 14, Gemini 4 Carbon in the high reasoning setting scores 85.4 percent with a 3.0 margin of error, Claude Fable 5.2 in max scores 82.8 with 5.3, and GPT-6.1 Sol in max scores 80.6 with 2.7. On the ArXivMath track Carbon's overall result is 94.28 percent with a 3.64 margin, first among eight systems, and its September 2026 ArXivMath run reached 96.23 percent (per the MathArena model page).

Accuracy has a price tag. With the reasoning settings used in testing, an expected Carbon run costs about 15.70 dollars against roughly 40 dollars for Fable 5.2 and about 1 dollar for GPT-6.1 Sol — per The Decoder, roughly 15 and 40 times the cost respectively. The MathArena listing for Carbon shows release date November 10, a 1,048,576-token maximum output window and current pricing of 2 dollars per million input tokens and 10 per million output tokens.

Both challengers are known quantities: Carbon is the model Google opened to all paying developers on November 10, and Fable 5.2 launched on November 9 — so this update is about measured results, not new products. MathArena publishes accuracy with 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 every model — so comparisons come with uncertainty ranges rather than single-point scores, a useful format given how close 85.4 and 82.8 actually are.

IMPACT
For automation teams this is a free decision aid: an independent benchmark refreshes without any vendor announcement, and the math track is a reasonable proxy for the multi-step analytical work agents do on spreadsheets, filings and research. The leaderboard also prices accuracy: the top result costs an order of magnitude more per run than the value option at 80.6 percent. Practical step: keep a benchmark-informed shortlist, but measure your own tasks — an automation budget feels the cost of a completed run, not a leaderboard point.

Sources: matharena.ai — the-decoder.com