The Overtaking
For the first time in a while, an OpenAI model is not leading the MathArena mathematical benchmark ranking. According to The Decoder, Google’s Gemini 4 Carbon has knocked GPT-6.1 Sol off the top, while Anthropic’s Claude Fable 5.2 placed second. According to the source, the result places the two models among the most capable currently available to the public for mathematics tasks.
The Decoder notes that OpenAI’s models—which were recently credited with solving complex equations such as the Navier-Stokes equations—are not yet publicly accessible. The distinction is relevant: the comparison concerns models actually available, not announced or claimed capabilities.
Behavior and Limitations Observed in the Tests
The most frequent weak point of Gemini 4 Carbon, writes The Decoder, is exceeding the time limit. The model, however, almost never gives up on completing a task: a trait that the source describes as open to positive or negative interpretation. In one case, the article reports, it started a script that remained running for more than eleven hours without checking its progress.
These are observations attributed to the source and limited to the tests carried out, not a general assessment of the model’s reliability. The material does not clarify whether the script produced a useful result or simply kept running: an element that remains to be verified.
The Price of First Place
The performance advantage has an economic cost. With the reasoning settings used in the tests, The Decoder reports, Gemini 4 Carbon and Claude Fable 5.2 cost approximately 15 and 40 times more than GPT-6.1 Sol, respectively. The comparison is therefore twofold: on one hand, leadership in the mathematical tests; on the other, a significantly higher outlay than the OpenAI model.
The material does not indicate the scores obtained by the models, the exact composition of MathArena, or the test configurations, so it is not possible to establish how much of the cost gap is attributable solely to the reasoning settings. It also remains unclear whether the costs observed in the tests correspond to those of ordinary use. In short, the news signals a change at the top in a specific benchmark and an economic trade-off documented by the source, not an overall judgment on one model’s superiority.
