Someone just dropped a powerful new AI model onto OpenRouter’s API platform and walked away without leaving a business card. Ox Alpha, listed under the provider name “Stealth,” launched on August 20 with a 1,048,576-token context window, multimodal reasoning across text, image, and video inputs, and a price tag of exactly zero dollars for its one-week preview period.
The model’s anonymous debut has sent the developer community into full detective mode, with tokenizer fingerprinting and API signature matching pointing strongly toward a familiar suspect: China’s Zhipu AI, now known as Z.ai.
What Ox Alpha actually does
A million-token context window is the headline feature, and it’s worth understanding what that means in practice. Most large language models can process somewhere between 8,000 and 128,000 tokens at once. A million tokens is roughly equivalent to feeding the model an entire codebase, or several full-length novels, in a single prompt.
That kind of capacity makes Ox Alpha particularly suited for long-horizon software engineering tasks, where an AI agent needs to understand vast amounts of existing code before making changes. Early testers have already noted its seamless integration into coding agents and editors.
Initial benchmarks from the community suggest throughput of approximately 29 tokens per second, which is respectable for a model handling that much context. The model also supports tool and function calling alongside structured JSON output, the kind of capabilities that make it immediately useful for production workflows rather than just chat-style interactions.
For the preview period, the provider has stated that no prompts will be used for training.
The detective work
Ox Alpha marks the fifth time an anonymous AI model linked to Chinese labs has appeared on the OpenRouter platform.
The strongest evidence points to Ox Alpha being a variant of GLM-5.3, part of the GLM family developed by Zhipu AI. Tokenizer fingerprinting, which analyzes how a model breaks text into tokens, essentially creates a signature that’s hard to disguise. API response patterns add another layer of identification. Both point toward the GLM lineage.
A secondary theory involves Xiaomi’s MiMo team, which has used similar stealth release strategies in the past. Previous anonymous drops on OpenRouter, including GLM-5 and MiMo-V2-Pro, were eventually claimed by Chinese labs after the initial preview period ended.
No architecture details, formal benchmarks, or post-preview pricing have been disclosed. The model’s listing claims a servicing capacity of over 100 trillion tokens per day.