Gemini 4 Carbon tops MathArena but costs far more to run
Google’s Gemini 4 Carbon has topped the MathArena math benchmark, becoming the first non-OpenAI model in some time to claim the top spot, according to the-decoder.com. The model displaced GPT-6.1 Sol, with Anthropic’s Claude Fable 5.2 finishing second, positioning Carbon as one of the most capable publicly available models for math tasks.
The outlet notes that OpenAI models reportedly used to solve demanding equations such as the Navier-Stokes equations are not yet publicly available, leaving Carbon’s public benchmark win notable for now.
According to the-decoder.com, Gemini 4 Carbon’s most frequent failure mode is running past allotted time limits rather than producing wrong answers. The model also rarely abandons a task, a trait the outlet describes as cutting both ways: in one instance Carbon launched a script that ran for more than eleven hours without checking on its own progress.
That persistence and capability come at a cost. Under the reasoning settings used in testing, Carbon costs roughly 15 times as much to run as GPT-6.1 Sol, while Claude Fable 5.2 costs about 40 times as much, the-decoder.com reported.
The results add to a fast-moving competitive picture among frontier labs, with Google, OpenAI and Anthropic each pushing specialized reasoning models that trade higher compute costs for stronger performance on demanding technical benchmarks. The price gap highlighted by the-decoder.com underscores a growing tension in the field between raw capability gains and the economics of deploying increasingly expensive reasoning-heavy models at scale.
Based on reporting by the-decoder.com.
