THE CRUNCH
A non-OpenAI model has reclaimed the MathArena maths benchmark for the first time in a while: Google's Gemini 4 Carbon knocked GPT-6.1 Sol off the top spot, with Anthropic's Claude Fable 5.2 taking second. That puts Carbon among the strongest publicly available models for maths tasks, at least potentially.
The crown comes with caveats. Carbon's most common failure mode is running over the time limit, and it rarely abandons a task, which can be read as persistence or stubbornness. In one instance, it started a script that kept running for over eleven hours without any progress check.
Then there is the bill. Under the reasoning settings used in the tests, Carbon cost roughly 15 times as much as GPT-6.1 Sol, while Fable 5.2 ran to about 40 times as much. OpenAI's own models, which have allegedly solved demanding equations such as the Navier-Stokes equations, are not publicly available yet.