Gemini 4 Carbon tops MathArena, ahead of Claude Fable 5.2 and GPT-6.1 Sol, at a much higher cost
On the MathArena benchmark, Google’s Gemini 4 Carbon scored 85.4%, ahead of Anthropic’s Claude Fable 5.2 at 82.8% and OpenAI’s GPT-6.1 Sol at 80.6%, according to The Decoder, citing MathArena’s Jasper Dekoninck. Carbon cost about $15.70 per run versus $1.07 for GPT-6.1 Sol, roughly 15 times more; Fable 5.2 cost $40.20.
The DecoderOriginally published 1 min
Why it matters
It is the first time in a while that an OpenAI model does not lead this benchmark, and the top two models are both publicly available. Margins of error overlap among the leaders, and Carbon’s most common failure was exceeding the time limit.
The best open model, Kimi K3.1 Pro, scored 54.3%.
For business & IT
Benchmark rank is not cost-efficiency: compare score per dollar on your own tasks before switching models.
OpenAI says it is expanding its ChatGPT for Academic Researchers program to give 200,000 researchers worldwide free access to its frontier models. In a post on X, the company also said a study of more than 1,800 participants in the program found scientists reporting nearly double their productivity with AI.
OpenAI published a post on a candidate nonperturbative lattice definition of the Standard Model of particle physics. The company says its GPT-6 Astra model helped turn the idea into a detailed proof. We could not read the full post, so the physics and the model’s exact role are not independently confirmed.
According to The Decoder, Anthropic’s Claude Fable 5.2 creates pictures with Python programs that drive a simulated paint engine, rather than computing pixels like diffusion models. Anthropic employee Jake Eaton says one photorealistic parking-garage scene took roughly 700,000 strokes. The model uses no external image model or traced source images.
Meituan has released the weights of LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model for coding and agent tasks, under an MIT license, The Decoder reports. By Meituan’s own tests it scores 88.4 on Terminal Bench v2.1, just behind Claude Opus 5. Its sample deployment runs on 194 Huawei Ascend chips and the guide does not mention Nvidia GPUs.
The Decoder
With your permission, we use Google Analytics to understand which stories are read. No advertising, and your data is never sold. Privacy policy