Mystery model Ox Alpha traced to Z.ai

A free model called Ox Alpha appeared on OpenRouter on Thursday, August 20. Its listing named no developer, saying only that an anonymous third party was operating it for this preview, and described it as a reasoning model built for coding, sustained agentic work, and production use, TechCrunch reported August 23. Stripe CEO Patrick Collison, whose company is acquiring OpenRouter, called it “very impressive” on X, but for several days neither OpenRouter nor any lab confirmed who trained it, and speculation ran in several directions without converging, including guesses at Z.ai’s GLM line and an unreleased build of Microsoft’s MAI.

Update, August 26: Z.ai confirmed it built Ox Alpha, Bloomberg reported, identifying it as the newest model in Z.ai’s GLM line. Z.ai described it as designed for coding, sustained agentic work, and production workloads, including long-horizon software engineering and workflows that combine text with visual context, and said open weights would follow the same day. Ox Alpha was already topping usage and benchmark leaderboards on OpenRouter before the reveal, and Bloomberg frames the release as part of a wider pattern of cheaper Chinese open-weight models, including Z.ai’s recent GLM-5.3, competing directly with frontier labs like OpenAI and Anthropic on capability at a fraction of the price.

Update, August 27: Ox Alpha’s formal name is GLM-5.3-Flash. It’s a mixture-of-experts model with 320 billion total parameters and 18 billion active per token, accepting up to 1 million tokens of text, images, and video and returning responses up to 131,072 tokens, trained on a 30-trillion-token dataset. Z.ai ran it head to head with three other frontier models, Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash, and says it came out on top on the GDPval-AA v2 benchmark and placed second on AutomationBench, and says it costs roughly one-tenth as much to run as Z.ai’s prior-generation model, all figures from the model’s own maker that are worth checking against independent benchmarks before a production decision.

What it means for you

Anonymous stealth-model previews on OpenRouter and similar routers have become a routine way for labs to gather real-world usage signal before a named launch, and this one resolved the way most of them eventually do: to a real, accountable lab rather than an unknown operator. Now that it carries Z.ai’s name and is shipping as open weights, the practical calculus changes from “don’t route anything sensitive through an anonymous model” to the same routing and provenance questions you’d apply to any other Chinese open-weight release, cost, hosting terms, and whether its licensing and data-handling suit your use case, rather than a blanket no. Worth reading with our take on anonymous stealth-model risk if you tried it during the anonymous window and are now deciding whether to keep it in a real pipeline.