Google’s Gemini 4 Carbon now leads MathArena, the maths benchmark run by researchers at ETH Zurich and INSAIT. Claude Fable 5.2 is second, and GPT-6.1 Sol, which held the top place before, drops to third.
The results went up on 13 November. MathArena’s Jasper Dekoninck wrote that this is the first time in a very long while that a model from outside OpenAI is number one.
The higher scores are not free. Under the reasoning settings MathArena used, both new leaders cost significantly more to run than the model they replaced, and the gap is widest on its BrokenArXiv test. Carbon was so expensive that, as the team said in its announcement thread, it could only be run on one of the Lean tests, ArXivLean, where it did very well. They plan to add it to LeanLean later.
The team also logged how each model fails. Gemini mostly loses points by running past the time limit, because it keeps working instead of stopping. One run launched a script that went on for over 11 hours with no check on it. Fable’s usual problem is a bug in extracting its reasoning, and MathArena marks it down for that. The ArXivMath table for October is below: it lists each model’s accuracy, what the run cost and how many tokens it used, with a coloured cell for every problem, so a reader can see where a model slipped and what it paid to get there. MathArena’s own site holds the full results for every other test.
Gemini 4 Carbon opened to paying developers and Ultra subscribers on 10 November.
