Gemini 4 Carbon Tops MathArena — the First Non-OpenAI Model to Take the Math Crown
For a long stretch, the top of the MathArena leaderboard had a familiar look: an OpenAI model sitting in first place. That changed this week. Google’s Gemini 4 Carbon has taken the No. 1 spot on MathArena, posting 85.4% expected performance with its reasoning set to “high”. It’s the first time in a long while that a model from outside OpenAI has held the math crown.
The result surfaced on Nov 13, when benchmark operator Jasper Dekoninck announced it on X. It didn’t stay a niche post for long. Sundar Pichai and Demis Hassabis both reposted it, and that tells you how much weight Google puts on a win like this one.
One detail matters here: where the news came from. This wasn’t a Google press release or a launch-day blog post. The word came from the benchmark’s own operator, and the numbers are MathArena’s, published on its own leaderboard for anyone to check. Wins announced that way tend to carry a bit more weight than self-reported scores.
The numbers behind the crown
MathArena measures how well AI models handle hard mathematical reasoning, and its leaderboard has become one of the scoreboards the big labs watch closely. Carbon’s 85.4% puts clear air between it and the two models it had to beat:
- Gemini 4 Carbon — 85.4%
- Claude Fable 5.2 — 82.8%
- GPT-6.1 Sol — 80.6%
A 2.6-point gap over second place isn’t a photo finish. And the field behind it is strong. Anthropic’s entry is no slouch — we covered the Claude Fable 5.2 and Mythos 5.2 launch earlier this month, and hard reasoning was one of the areas it pushed on. Sol, meanwhile, has been the reference point at the top of this board, the score everyone else measured themselves against.
So what does “expected performance” actually mean in plain terms? Think of it as the score the benchmark expects the model to repeat across the problem set, not one lucky run on a good day. Consistency counts. A model that cracks a brutal problem once and fumbles it the next two times doesn’t get the same credit as one that solves it every time.
The catch is the bill
There’s a catch, and it’s a big one: cost. According to the benchmark coverage, Carbon burned through roughly $15.70 per problem to post that score. Sol did its work for about $1.07 per problem. That’s not a rounding difference — it’s close to fifteen times the spend for under five extra points of performance.
Does that make the result meaningless? Not at all. Leaderboards measure raw capability, and a capability headline like this one matters. It shows what the model can do when it’s allowed to think as hard as it wants, with the meter running.
But anyone planning to run a reasoning model at real scale reads the cost column first. A model that’s brilliant and expensive faces a very different job than one that’s nearly as good and cheap to run. For researchers chasing the hardest problems, Carbon’s approach makes sense. For a product team serving millions of queries a day, the maths of the invoice matters just as much as the maths on the benchmark.
Why Google needed this win
Google has spent the past year arguing that Gemini belongs at the very top of the frontier conversation, and the Gemini lineup on DeepMind’s site keeps growing to back that claim up. Arguments are cheap, though. Being the first non-OpenAI model on top of MathArena in a long while gives the claim a concrete receipt — a public scoreboard, a date, and a margin nobody has to take Google’s word for.
It also lands in a busy stretch for Google’s engineering story. The company recently open-sourced XSE, the serving engine behind six million queries a second. Different story, same theme: Google wants to be seen setting the pace in this race, not chasing it.
What happens next
Leaderboard leads rarely last long, and nobody at Google will be pretending otherwise. The obvious questions now are whether Carbon can hold the top spot once its rivals respond, and whether Google can drag that cost per problem down without handing the points back. Reasoning set to “high” is clearly where the score came from, so efficiency is the next battleground.
For now, the board says what it says. The math crown sits with Google — and for the first time in a long while, it isn’t wearing OpenAI colours.