Google’s model sits at the top of the MathArena benchmark, ahead of GPT-6.1 Sol and Claude Fable 5.2. This advance is accompanied by frequent deadline overruns and a significantly higher cost with the reasoning parameters tested, while some OpenAI models mentioned are not yet publicly available.
Time overruns and rising invoice under the tested settings
The most common failure mode of Gemini 4 Carbon is exceeding the time limit. The model rarely abandons a task, which can be seen as an advantage or a disadvantage. In one example, a script ran for over eleven hours without checking its progress. With the reasoning parameters used in testing, Gemini 4 Carbon costs about 15 times more than GPT-6.1 Sol, and Claude Fable 5.2 about 40 times more.
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Limited public access for some OpenAI models
Reportedly, models developed by OpenAI have recently been able to solve complex equations such as the Navier-Stokes equations, however these models are not yet made available to the public. MathArena constitutes in this context a point of comparison for measuring the mathematical skills of accessible models.
Current leaderboard rankings and announced milestones
Gemini 4 Carbon, developed by Google, became the first non-OpenAI model to take the lead in the MathArena benchmark in some time, knocking GPT-6.1 Sol out of the top spot. Claude Fable 5.2 takes second place, and Gemini 4 Carbon is among the best performing public models for mathematical tasks. Expected performance is set for November 10, 2026 for Gemini 4 Carbon (top), November 9, 2026 for Claude-Fable-5.2 (max), September 29, 2026 for GPT-6.1 Sol (max) and October 23, 2026 for Kimi K3.1 Pro (max).

