Anonymous AI Model 'Ox Alpha' Crushes Coding Benchmarks, Sparking a Cross-Country Guessing Game

Its specifications are equally aggressive: a 1,048,600-token context window and up to 131,000 tokens of output per call, with text, image and video input and native support for tool calling, function calling and structured JSON output. In the fuller 113-task DeepSWE evaluation, Ox Alpha scored 63, placing it close to Claude Opus 4.8.

A cross-border guessing game followed. Developers used “technical fingerprinting” to infer the source: a tokenizer test found a constant 75-token difference from Zhipu AI’s GLM-5.3, a video-encoder test matched GLM-5V-Turbo token-by-token, and an API error format echoed Zhipu’s services. Many now put 99 percent confidence on Ox Alpha being a pre-release Zhipu GLM-5.x model. Google DeepMind researchers added to the suspense with guarded hints about Gemini, though no company has claimed the model.

Ox Alpha is not the first anonymous launch. Earlier stealth models on OpenRouter all turned out to be Chinese: Pony Alpha was Zhipu’s GLM-5, Hunter Alpha was Xiaomi’s MiMo-V2-Pro, Elephant Alpha was Ant’s Inclusion AI, and Owl Alpha was Meituan’s Long Cat-2.0.

Why post anonymously? Stripping away brand bias lets developers judge a model purely on output, and quietly seeding free, top-performing models has become a recurring strategy for Chinese model makers — the next move in the compute business.

Tags