Gemini 4 Carbon tops MathArena with higher computing costs
Gemini 4 Carbon leads the MathArena mathematics assessment at the tested reasoning setting, ahead of Claude Fable and GPT Sol, according to results published by The Decoder. Its expected cost is also substantially higher than Sol’s. Time limits are Carbon’s most common failure, and one attempt left a script running for more than eleven hours without checking its progress.
Artificial Intelligence··Midday
Carbon takes the mathematics lead
Google’s Gemini 4 Carbon leads the MathArena assessment of AI mathematics performance, according to The Decoder’s November 14 account. At high reasoning effort its expected score is 85.4 per cent. Claude Fable 5.2 at maximum effort follows with 82.8 per cent, ahead of GPT-6.1 Sol at maximum effort on 80.6 per cent. The comparison concerns the models at those specified settings.[1]
Expected cost rises with the leader
The same assessment puts Carbon’s expected cost at 15.70 dollars, compared with 1.07 dollars for Sol and 40.20 dollars for Fable. Carbon’s cost is approximately 15 times Sol’s under the reported settings. Moonshot AI’s Kimi K3.1 Pro is listed as the strongest open model, with expected performance of 54.3 per cent at maximum effort and an expected cost of 3.70 dollars.[1]
Long attempts run past the limit
Carbon most often fails by running out of allotted time in this assessment. Tasks are almost never abandoned by the model. One attempt involved a script running for more than eleven hours with no check on its progress. The model arrived on November 10; Fable was released on November 9, and Sol on September 29.[1]
Related columns
For more information on this topic, you can read the related columns.