Gemini 4 Carbon Tops MathArena, Costs 15x More

Google's Gemini 4 Carbon now sits atop MathArena, the benchmark tracking how well AI models handle demanding math problems. It pushed GPT-6.1 Sol out of first place, with Anthropic's Claude Fable 5.2 landing in second. For a stretch, OpenAI models had the top spot locked down, so a non-OpenAI model winning marks a change worth noting. OpenAI's own frontier systems, reportedly capable of cracking problems as thorny as the Navier-Stokes equations, remain unreleased, which leaves the public leaderboard open to competitors willing to spend.

And spending is the operative word here. Topping a math benchmark isn't just about raw capability. It also comes down to how a model behaves when a problem resists a quick answer: whether it checks its own work, when it decides to quit, and how much compute it's willing to burn chasing a solution.

Gemini 4 Carbon's ranking raises a practical question for anyone choosing between these systems, because the gap between "best" and "most expensive" turns out to be substantial. What that gap looks like in practice, and how Carbon's problem-solving habits factor in, is where the numbers get interesting.