Gemini 4 Carbon from Google has taken the top spot on the MathArena benchmark, displacing OpenAI’s GPT-6.1 Sol. This achievement marks the first time a non-OpenAI model has led this specific test in some time. Claude Fable 5.2 from Anthropic finished in second place. The leading models are not yet publicly available for general use, which limits direct comparison with the open weights community.
The superior performance comes with a significant financial penalty. When using the reasoning settings required for the test, Gemini 4 Carbon costs roughly fifteen times as much as GPT-6.1 Sol. The model also exhibits a distinct failure mode involving excessive runtime. In one instance, it executed a script for over eleven hours without monitoring progress. This relentless approach to problem solving drives up operational costs while potentially wasting compute resources on unproductive loops.
- Gemini 4 Carbon scored 85.4 per cent accuracy on the benchmark.
- Its cost per query was estimated at fifteen dollars and seventy cents.
- The model failed to stop running a specific task after eleven hours.
