The anonymous AI model that topped OpenRouter's usage charts within days is a new GLM-family release from China's Z.AI, which confirmed the launch and will open-source its weights.
The mystery AI model that swept to the top of OpenRouter's usage charts is a new GLM-family release from China's Z.AI, which confirmed the launch Wednesday and said it will release the model's weights tonight. The Beijing-based company, also known as Zhipu, told Bloomberg that Ox Alpha is the latest iteration of its GLM series, ending a week of speculation after the model appeared on the marketplace Aug. 20 with no creator attached.
"It's very impressive," Patrick Collison, chief executive of Stripe Inc., which is acquiring OpenRouter, said on X. The endorsement from a prominent tech executive helped push Ox Alpha into wider conversation within hours of its debut.
Ox Alpha's usage more than doubled DeepSeek's on OpenRouter's leaderboard, making it the biggest launch in the marketplace's history, according to the company. The reasoning model carries a 1,048,600-token context window, supports native tool calling and accepts text, image and video input. On the DeepSWE benchmark, which scores how often a coding agent resolves real GitHub issues on its first attempt, Ox Alpha passed 63 percent of 113 tasks — roughly level with OpenAI's GPT-5.6 Sol, though a viral 10-task sample had earlier shown it at 80 percent. OpenCode said its infrastructure could handle 100 trillion tokens a day during the free window.
The open-weights strategy lets developers build applications on top of the model, intensifying pressure on US rivals as Chinese labs ship frontier-class systems at lower cost. Z.AI is offering Ox Alpha free for one week and has not disclosed pricing afterward. The code name was inspired by the Chinese film "Niu Lai," or "Ox Comes," the company said.
Open weights and the anonymous-launch playbook
Z.AI's decision to release the weights — the parameters that guide how an AI system processes information — follows a pattern Chinese labs have used all year to test frontier products before a formal announcement. Alibaba and Xiaomi released models anonymously this year, and Z.AI itself unmasked its Pony Alpha as GLM-5 in February after five days. Xiaomi's Hunter Alpha turned out to be MiMo-V2-Pro in March, and Meituan's Owl Alpha ran for two months before being confirmed as LongCat-2.0 in June.
Before Z.AI confirmed its involvement, developers had already fingerprinted Ox Alpha as a GLM product. Independent analysis matched its tokenizer to GLM-5.3 on every normalized test and its video-token spend to GLM-5V-Turbo exactly, while sharing GLM's distinctive error codes and audio-rejection behavior. GLM-5.3 launched Aug. 14 as a text-only model, suggesting Ox Alpha is an unreleased multimodal upgrade to that line.
What it means for the AI race
Ox Alpha arrives as Chinese companies keep releasing benchmark-topping models at a fraction of US development costs. Z.AI introduced GLM-5.3 earlier this month, which it said can rival Anthropic's Fable 5 on some benchmarks, while DeepSeek and Moonshot AI have drawn global attention with their own releases. The company, which listed publicly in January, will issue its first detailed earnings report Monday, covering its first six months as a public company.
For investors, the open-weights model threatens to compress pricing power across the AI stack. If Ox Alpha sustains its usage lead after the free window closes, it could pull developer demand away from paid frontier models from OpenAI and Anthropic, whose revenue depends on API usage. Z.AI has not disclosed the cost of running Ox Alpha or its post-trial pricing, leaving the economics of the challenge unquantified for now.
This article is for informational purposes only and does not constitute investment advice.