Gemini 4 Carbon tops MathArena at 85.4%, ahead of Claude Fable 5.2 and GPT-6.1 Sol

Google's Gemini 4 Carbon has taken first place on the MathArena math benchmark with an expected performance of 85.4% (±3.0%), according to results shared by Jasper Dekoninck on X and reported by The Decoder. It displaces OpenAI's GPT-6.1 Sol, making Carbon the first non-OpenAI model in a while to top the leaderboard. Anthropic's Claude Fable 5.2 placed second.

The price of that lead is high. With the reasoning settings used in testing, Carbon and Fable 5.2 cost roughly 15 and 40 times as much as GPT-6.1 Sol, respectively.

MathArena results

Rank	Model	Provider	Release date	Expected performance	Expected cost
#1 overall	Gemini 4 Carbon (high)	Google	Nov 10, 2026	85.4% ±3.0%	$15.70 ±$1.33
#2 overall	Claude Fable 5.2 (max)	Anthropic	Nov 9, 2026	82.8% ±5.3%	$40.20 ±$2.59
#3 overall	GPT-6.1 Sol (max)	OpenAI	Sep 29, 2026	80.6% ±2.7%	$1.07 ±$0.018
Best open model	Kimi K3.1 Pro (max)	Moonshot AI	Oct 23, 2026	54.3% ±5.8%	$3.70 ±$0.27

The parenthetical in each model name is the reasoning setting used in testing. Per-token API pricing and context window sizes were not part of the reported results.

Failure modes

Gemini 4 Carbon's most common failure mode is exceeding the time limit. The model also almost never gives up on a task, a trait that can be read as either positive or negative. In one case, it launched a script that ran for more than eleven hours without checking on its progress.

Availability caveat

The report notes that OpenAI's models, which recently allegedly solved demanding equations such as the Navier-Stokes equations, are not publicly available yet. On that basis, Gemini 4 Carbon is potentially one of the most capable publicly available models for math tasks.

What this means

On MathArena, Gemini 4 Carbon has the highest expected performance of the models listed, and Claude Fable 5.2 follows it. GPT-6.1 Sol has the lowest expected cost of the three at $1.07, against $15.70 for Carbon and $40.20 for Fable 5.2. Teams choosing a model for math-heavy work will need to weigh accuracy against cost.

Carbon's tendency to hit time limits, and its habit of persisting on a task without checking progress, matter for anyone running it in automated pipelines. The eleven-hour script is an example of that behavior. The strongest open model on the list, Kimi K3.1 Pro (max), scores 54.3% (±5.8%) at an expected cost of $3.70.