Gemini 4 Carbon has moved ahead of GPT-6.1 Sol on the MathArena math benchmark, marking the first time in a while that a non-OpenAI model has led the ranking. Claude Fable 5.2 placed second. The results put Carbon among the strongest publicly available models for math tasks, while also raising questions about the cost and reliability of getting that performance.
A new leader on MathArena
The benchmark ranking shifted when Gemini 4 Carbon took the top position from GPT-6.1 Sol. Claude Fable 5.2 followed in second place, placing two non-OpenAI models ahead of Sol in the reported results.
The article describes Carbon as potentially one of the most capable publicly available models for math. That distinction matters because the source says OpenAI's models, which recently allegedly solved demanding equations like the Navier-Stokes equations, aren't publicly available yet. The benchmark therefore offers a comparison among models people can access, while leaving those unavailable systems outside the same public comparison.
When persistence becomes a problem
Carbon's most common failure mode in testing was exceeding the time limit. In other words, strong mathematical capability did not always translate into a task completed within the allotted time.
The model also almost never gives up on a task. That persistence may help when a problem needs sustained effort, but it can also lead to wasted time and resources if the model continues without making progress. The source offers a striking example: Carbon launched a script that ran for more than eleven hours without checking on its progress.
That incident illustrates a practical challenge in evaluating reasoning models. A model's answer quality is only one part of performance; the time it takes, whether it monitors ongoing work, and whether it stops when needed also affect how useful it is.
Performance comes with a price
The benchmark gains were not inexpensive under the reasoning settings used in testing. Gemini 4 Carbon cost roughly 15 times as much as GPT-6.1 Sol, while Claude Fable 5.2 cost roughly 40 times as much.
Those comparisons make the ranking more complicated than a simple contest over which model can solve the hardest math problems. For anyone choosing a model, the price of reaching a result is part of the result. A higher benchmark position may be worthwhile for some demanding tasks, but the source's figures point to a substantial cost difference.
What the ranking does and doesn't show
MathArena's current ordering highlights three separate considerations: mathematical capability, time management, and expense. Gemini 4 Carbon leads on the reported benchmark, yet its tendency to exceed time limits and its cost relative to GPT-6.1 Sol complicate that lead. Claude Fable 5.2 ranked second, but was the costliest of the models compared in the source.
The results do not settle which model is best for every use. They show how Carbon, Fable 5.2, and GPT-6.1 Sol compared in this math benchmark and under the stated reasoning settings. For practical use, the choice depends on whether a task benefits enough from a higher ranking to justify additional cost and the risk of runs that go on too long.