The Decoder reported on 14 November 2026 that Google's Gemini 4 Carbon now tops MathArena, a math benchmark. Its table gives Carbon 85.4% ±3.0%. Anthropic's Claude Fable 5.2 follows at 82.8% ±5.3%, and OpenAI's GPT-6.1 Sol at 80.6% ±2.7%. Those margins overlap, so the order is not clear-cut. Carbon ran at the 'high' reasoning setting, which controls how much a model thinks. The other two ran at 'max'.
The Decoder's table lists expected costs of $15.70 for Carbon, $40.20 for Fable 5.2 and $1.07 for GPT-6.1 Sol. The outlet puts Carbon at roughly 15 times Sol's cost and Fable 5.2 at roughly 40 times. It says Carbon most often fails by running past the time limit, and it almost never gives up. In one case, a script ran for over eleven hours with no progress check. The report does not say what that run cost, or why the prices differ so much.
A roughly 15-fold price buys a lead that sits inside the error range, and the models ran at unequal settings. A model that keeps working on its own is a cost question as well as a quality one. Buyers can ask what each solved problem costs, what time and spending caps exist, and whether their work is really math-heavy.