Developers can test Ox Alpha for free this week, a capable new model with a context window of just over a million tokens that is built for coding, long-horizon agent work, and production use. It arrived on OpenRouter as a stealth release from a third-party provider that has not identified itself. The provider retains every prompt and completion it receives, which is the detail that matters most before you send it real work.<\/p>\n
Ox Alpha appeared on OpenRouter last week as an anonymous release. 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. The scale is significant, and the model is positioned for coding, agent tasks that run over long sessions, and live production use.<\/p>\n
Early hands-on reactions have been positive. Developers who tried it rated it highly, and one senior technology executive who tested it called it "very impressive." The combination of a million-token context window and free access for a week has drawn a large amount of testing across the industry. A stealth launch of this kind is a recognized way to benchmark a model against real usage before a formal announcement.<\/p>\n
The provider has not said who it is, and the guessing has stayed inconclusive. One leading theory points to Z.ai, which previously tested a model anonymously under a different name before releasing GLM-5. A competing analysis of the model’s tokenizer suggests Microsoft’s MAI family instead. Confidence in every theory dropped over the following days as more people examined it, and no candidate has been confirmed. Somebody is paying for 100 trillion tokens a day of inference, and until they say who they are, the source of that spending remains unknown.<\/p>\n
OpenRouter’s own listing states that prompts and completions "are retained by the provider and are not used for training." That means whatever you send goes to a company that has not disclosed its identity, and that company keeps it. For a personal experiment this is a manageable trade. For a work project that includes confidential material, it changes the calculation.<\/p>\n
Data protection law requires a contract with a named data processor and an assessment of where data travels. Neither is possible when the counterparty is anonymous. The timing adds pressure: the AI Act’s transparency obligations took effect on 2 August, with penalties reaching €15mn or 3% of global turnover, in a framework built on knowing which provider is responsible for what. An unnamed provider that retains user input sits outside what that framework allows a business to sign off on.<\/p>\n
None of this makes the model itself weak. Open-weight releases have closed the capability gap quickly, and testing a model quietly before announcing it is a legitimate practice. The practical cost of free access here is information about who receives your data. A sensible approach for any business handling sensitive material is to try Ox Alpha only with content that does not matter until the provider is named.<\/p>\n
Ox Alpha is an AI model released anonymously on OpenRouter, free to use with a context window of just over a million tokens. It is positioned for coding, long-horizon agent work, and production use, and OpenCode said it would be free for a week with near unlimited usage.<\/p>\n
Yes. OpenRouter’s listing states that prompts and completions are retained by the provider and are not used for training. Because the provider has not identified itself, businesses cannot sign the named-processor contract that data protection law requires.<\/p>\n
The provider is unidentified. One theory points to Z.ai, which previously tested a model anonymously before releasing GLM-5, while an analysis of the tokenizer suggests Microsoft’s MAI family. No candidate has been confirmed.<\/p>\n