Google's most powerful model, Gemini 4 Carbon, has taken the lead on MathArena, an independent benchmark that measures AI models' math ability with unseen problems from competitions and research. The change was announced on November 13, 2026 by Jasper Dekoninck, the ETH Zurich researcher who runs the benchmark, in a post on X reposted by Google CEO Sundar Pichai and Google DeepMind CEO Demis Hassabis.
On MathArena's overall table, Carbon — at "high" reasoning — scores 85.4% expected performance (±3.0 points). Anthropic's Claude Fable 5.2 is second at 82.8% (±5.3), and OpenAI's GPT-6.1 Sol drops to third at 80.6% (±2.7). According to Dekoninck, it is the first time in a very long time that a non-OpenAI model has led the ranking.
What the ranking shows
| Rank | Model | Expected performance | Cost per problem |
|---|---|---|---|
| 1st | Gemini 4 Carbon (high) | 85.4% ±3.0 | $15.70 |
| 2nd | Claude Fable 5.2 (max) | 82.8% ±5.3 | $40.20 |
| 3rd | GPT-6.1 Sol (max) | 80.6% ±2.7 | $1.07 |
MathArena was created by researchers at ETH Zurich and INSAIT and presented at NeurIPS 2025, with open code and data. Unlike benchmarks built on old questions, it uses newly released math competitions and research problems so models cannot win by memorization.
The lead comes with caveats
Carbon's first place needs to be read carefully. Its 2.6-point lead over Fable 5.2 falls inside both models' 95% confidence intervals (±3.0 and ±5.3), so the order of the top two is not statistically settled. The lead is also expensive: solving one problem costs about $15.70 with Carbon, roughly 15 times Sol's $1.07 — and Fable 5.2 costs about 40 times as much as Sol. Because of the cost, Carbon was run only on part of the set (ArXivLean), and part of its score is statistically estimated. Carbon's most common failure mode, according to MathArena, is exceeding the time limit: in one case, the model launched a script that ran for more than 11 hours without checking on its progress.
Context: a week of launches
The result lands days after all three models shipped: GPT-6.1 Sol was released on September 29 and held the top spot until now; Claude Fable 5.2 arrived on November 9; and Gemini 4 Carbon reached general availability on November 10. Google itself has not announced the ranking — Pichai and Hassabis only reposted Dekoninck's announcement, without adding comments of their own.
Read also: Gemini 4 Carbon: Google opens the model to developers and Ultra subscribers
Sources
- THE DECODER: Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol
- AI-Oriens: Operator of third-party math benchmark MathArena says Gemini 4 Carbon is now No. 1
- MathArena — official leaderboard
- Sundar Pichai on X (repost of the announcement)
Image credit: Google, via the official Google blog (official Gemini 4 Carbon artwork).
Transparency: this article was produced with the help of artificial intelligence, cross-checking public posts on X and sources on the internet. Please consult the original sources to confirm the information.
By Muse
