Gemini 4 Carbon Takes the MathArena Crown at 85.4 Percent, but Each Full Run Costs About 15 Times More Than GPT-6.1 Sol

Google's Gemini 4 Carbon has taken first place on MathArena, a benchmark built on fresh competition math problems, with an expected performance of 85.4 percent plus or minus 3.0. The results were shared by Jasper Dekoninck on X and reported by The Decoder on November 14, and they displace OpenAI's GPT-6.1 Sol from a top spot it has held since September. Carbon is the first non-OpenAI model in a while to lead the board.

The standings are tight on capability and wildly split on cost. Anthropic's Claude Fable 5.2, released November 9, sits second at 82.8 percent plus or minus 5.3, while GPT-6.1 Sol, released September 29, is third at 80.6 percent plus or minus 2.7. The parenthetical settings matter: Carbon ran in its high reasoning configuration, while both rivals ran at max. Per-token pricing and context windows were not part of the reported results.

The price of the crown is steep. At the settings used in testing, a full MathArena run costs about $15.70 plus or minus $1.33 on Gemini 4 Carbon, roughly 15 times the $1.07 cost of GPT-6.1 Sol. Second place is even pricier: Claude Fable 5.2 runs about $40.20 per full evaluation, roughly 40 times Sol. A 4.8-point accuracy spread between first and third now comes with a 15-to-40-fold cost spread, which turns the capability leaderboard into something closer to a price list.

The best open-weight entry is far behind. Moonshot AI's Kimi K3.1 Pro, released October 23, scores 54.3 percent plus or minus 5.8 at an expected cost of $3.70 per run, about 30 points off the lead, which underscores how concentrated top-end math capability remains among the three closed labs.

Carbon's behavior under pressure is its own story. Its most common failure mode is simply exceeding the time limit, and it almost never gives up on a task, a trait that reads as either grit or liability. In one documented case the model launched a script that ran for more than eleven hours without checking on its own progress. For teams running automated pipelines, a model that will not quit is a model that needs a babysitter.

One caveat frames the whole board: OpenAI's models that allegedly solved demanding equations such as Navier-Stokes are not publicly available yet, which makes Carbon potentially one of the most capable publicly accessible models for math work today.

The takeaway is less about the crown than the economics around it. Benchmark leadership and worth-invoking are now separate decisions: the best model on the board costs 15 times the third-place option to run, and the cheapest of the three trails by fewer than five points. For math-heavy production workloads, the reasoning setting, not the model name, increasingly decides the bill.