Developers now have a free coding model with a context window of just over a million tokens: Ox Alpha, which appeared on OpenRouter last Thursday. The model is positioned for coding, long-horizon agent work, and production use, and its provider has offered it free for a week with near unlimited usage. One detail shapes how you should use it: OpenRouter’s listing states that prompts and completions are retained by the provider, which has not said who it is.<\/p>\n
Ox Alpha launched as a stealth release from an anonymous third-party provider. It is free to use, with a context window of just over one million tokens, which is 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. A stealth launch is a normal way to benchmark a model against real workloads before an official announcement.<\/p>\n
Early reactions from developers have been positive, with the model described as very impressive and framed as ready for coding and production tasks. Serving 100 trillion tokens a day of inference points to a provider with significant infrastructure behind the release.<\/p>\n
The identity of the provider is the open question, and two theories have circulated. The leading one points to Z.ai, which previously tested its GLM-5 model anonymously under a different name. A competing analysis of the model’s tokenizer suggests Microsoft’s MAI family instead. Neither has been confirmed. Over the course of the weekend, confidence in every theory dropped, and observers described being less sure of the answer than they had been the night before.<\/p>\n
The release lands in a market that free, openly licensed models have already reshaped. Open-weight releases have narrowed the capability gap between models quickly, which is part of why a strong model can appear without a name attached and still be taken seriously.<\/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 the data. For casual testing this is a footnote. For work involving anything sensitive, it is the deciding factor, because you cannot evaluate a data handler you cannot name.<\/p>\n
For companies operating under European data protection law, an anonymous counterparty is a practical blocker. The law requires a contract with a named processor and an assessment of where data travels, and neither is possible when the other side is unidentified.<\/p>\n
The timing adds pressure. The AI Act’s transparency obligations took effect on 2 August, with penalties reaching €15mn or 3% of global turnover. That regime is built on knowing which provider is responsible for what, so a model with no named provider sits awkwardly against it. None of this makes Ox Alpha a poor model. It means the free access comes with a cost measured in information, and until the provider identifies itself, the safe approach is to test Ox Alpha with nothing that matters.<\/p>\n
Ox Alpha is a free AI model released anonymously on OpenRouter last Thursday, with a context window of just over one million tokens. It is positioned for coding, long-horizon agent work, and production use, and its provider had capacity for 100 trillion tokens a day.<\/p>\n
OpenRouter’s listing states that prompts and completions are retained by the provider and are not used for training. The provider that holds the data has not disclosed its identity.<\/p>\n
It has not been confirmed. One theory points to Z.ai, which previously tested GLM-5 anonymously under a different name, while an analysis of the tokenizer suggests Microsoft’s MAI family. The guessing has been inconclusive.<\/p>\n