Developers can now test Ox Alpha for free on OpenRouter, a model that ships with a context window of just over a million tokens and is positioned for coding, long-horizon agent work, and production use. The catch sits in the fine print: the provider behind it has not identified itself, and OpenRouter’s own listing states that prompts and completions are retained by that provider. Ox Alpha appeared last Thursday as a stealth release, free to use, from an anonymous third party.<\/p>\n
Ox Alpha is an anonymous model that surfaced on OpenRouter as a stealth release. It is free to use, and its context window runs to just over a million tokens, enough to hold large codebases and long agent sessions in a single prompt. The open-source agent OpenCode said the model would be free for a week with near unlimited usage, and that its provider had capacity for 100 trillion tokens a day.<\/p>\n
Early hands-on reactions have been strong. The model has drawn praise from developers who tried it, and it is being pitched for coding, extended agent tasks, and real production workloads rather than as a demo. That combination of scale, speed, and zero cost is why it spread quickly across the industry within days of going live.<\/p>\n
The identity remains an open question. One leading theory points to Z.ai, which previously tested GLM-5 anonymously under a different name. A competing analysis of the model’s tokenizer instead suggests Microsoft’s MAI family. By the weekend, confidence in either explanation had faded, and observers who tracked the debate ended up less certain than they had been a day earlier.<\/p>\n
Stealth launches like this are a recognized way to benchmark a model against the field before an official announcement. Open-weight releases have closed the capability gap faster than the safety one, so a strong unnamed model arriving without fanfare is now a normal event rather than a surprise.<\/p>\n
The more useful detail is not the name behind Ox Alpha but the terms attached to it. OpenRouter’s listing states that prompts and completions are retained by the provider and are not used for training. Read with a work project in mind, that means whatever you send goes to a company that has not said who it is, and that company keeps it.<\/p>\n
For European businesses, this is a practical blocker rather than a curiosity. Data protection law requires a contract with a named processor and an assessment of where data travels, and neither is possible when the counterparty is anonymous. There is no entity to sign an agreement with and no way to document the destination of the data.<\/p>\n
The regulatory timing adds weight. The AI Act’s transparency obligations took effect on 2 August, with penalties reaching €15mn or 3% of global turnover. The regime is built on knowing which provider is responsible for what, which is exactly the information an anonymous release withholds.<\/p>\n
None of this makes Ox Alpha a poor model. The capability is real, the context window is large, and the price is zero. The cost shows up as information instead of money. Someone is funding 100 trillion tokens a day of inference, and until that party names itself, the sensible approach for European teams is to test Ox Alpha with nothing that matters: no client data, no proprietary code, no confidential documents. Run it on throwaway inputs, measure the output quality, and hold real workloads back until the provider is known and a proper processing agreement is possible.<\/p>\n
Ox Alpha is an anonymous AI model released as a stealth launch on OpenRouter. It is free to use, has a context window of just over a million tokens, and is positioned for coding, long-horizon agent work, and production use.<\/p>\n
OpenRouter’s listing states that prompts and completions are retained by the anonymous provider, though not used for training. Because the provider is unnamed, European data protection law and the AI Act’s transparency rules make it impractical to use with sensitive data, so testing with non-critical inputs is the safer path.<\/p>\n
The provider has not identified itself. One theory points to Z.ai, which previously tested GLM-5 anonymously under another name, while an analysis of the tokenizer suggests Microsoft’s MAI family. Neither has been confirmed.<\/p>\n