Gemini 4 Carbon Just Took the Math Crown From OpenAI

The MathArena leaderboard has a new name at the top, and for once it isn't an OpenAI model.

Google's Gemini 4 Carbon scored 85.4% expected performance on MathArena with reasoning set to "high", edging out Claude Fable 5.2 (82.8%) and GPT-6.1 Sol (80.6%). Benchmark operator Jasper Dekoninck announced the result on X on Nov 13, and both Sundar Pichai and Demis Hassabis reposted it — a fair sign of how much this one matters internally at Google.

Why it stands out

MathArena scores models on hard mathematical reasoning, and "expected performance" rewards consistency across the problem set rather than one lucky solve. OpenAI models have held the No. 1 spot for a long stretch, so a 2.6-point lead for a rival is a genuine shift, not noise.

The expensive part

There's a catch hiding in the cost column. Per the benchmark coverage, Carbon spent roughly $15.70 per problem to get that score. Sol managed its 80.6% for about $1.07 per problem. Nearly fifteen times the spend for under five points of extra performance.

That doesn't erase the result — leaderboards measure what a model can do. But if you're planning to run reasoning models at scale, the cost per problem is the number that decides what actually ships.

What to watch

Can Carbon hold the top spot when the next rival models land? And can Google cut the cost without giving the points back? Reasoning at "high" is clearly where the score came from, so efficiency is where this fight goes next.

I broke down the full result, the leaderboard, and what it means for the frontier race here: Gemini 4 Carbon Tops MathArena