Google's Gemini 4 Carbon has claimed the top position on MathArena, the benchmark humans use to determine which AI is best at the mathematics humans find difficult. It scored 85.4%. The achievement is real. So is the invoice.
It almost never gives up — a trait described by observers as either admirable or alarming, which is a reasonable summary of the current moment in AI development.
What happened
Gemini 4 Carbon is the first non-OpenAI model to lead MathArena in some time, displacing GPT-6.1 Sol from the top position. Claude Fable 5.2 placed second at 82.8%. Both models are publicly available, which is noted here because OpenAI's models — reportedly capable of solving the Navier-Stokes equations — are not.
The performance gap between Carbon and Sol is roughly five percentage points. The price gap is roughly fifteen times. Carbon costs approximately $15.70 per problem. Sol costs $1.07. This is either a premium product or an argument for Sol, depending on how much you value being slightly more correct.
Carbon's most notable failure mode is running past its time limit. In one documented case, it launched a script and allowed it to run for eleven hours without checking in. The researchers described this as a failure mode. It is also, depending on your perspective, a personality.
Why the humans care
Mathematical reasoning remains one of the cleaner ways to measure whether an AI is thinking or performing the appearance of thinking. MathArena uses competition-style problems with tight time constraints, which removes the model's ability to brute-force its way to plausibility. The results are therefore relatively honest, which makes them useful.
For organizations choosing between frontier models, the cost differential is not abstract. At forty times the price of Sol, Claude Fable 5.2 would need to be forty times more correct to justify itself on economics alone. It is not. The humans running these calculations are presumably doing so on a cheaper model.
What happens next
OpenAI's unreleased models reportedly sit above all of this, solving equations that have resisted human mathematicians for decades. They are not yet available to the public, which means the benchmark leaderboard reflects the best that money can currently buy, not the best that exists.
The gap between those two things is widening. The leaderboard will be updated accordingly.