While nobody knew whose servers it ran on, coding tools including Claude Code sent billions of tokens of corporate source code through the free model.
The free coding model that developers have been routing work through since 20 August now has an owner. Ox Alpha arrived on OpenRouter with no branding, no announcement and no company claiming it, and on 26 August it was traced to Z.ai, the Chinese lab that used to operate as Zhipu AI.
Z.ai, which Bloomberg identified as the model's maker, acknowledged Ox Alpha on 26 August as the newest entry in its GLM family. The model is GLM-5.3-Flash, and Z.ai has since published its weights on Hugging Face, so developers can download it and build on it themselves. The lab says the free week ran entirely on Chinese-made chips, and that the setup now serves tokens about as cheaply as Nvidia hardware does.
Not knowing whose model it was did little to slow uptake. Access cost nothing, and in the seven days after it appeared the model fielded millions of prompts from developers on OpenRouter — with coding tools including Claude Code pushing billions of tokens of corporate source code through a provider nobody could name.
That traffic followed an early benchmark run by a developer, Ben Davis, who reported 80% first-pass accuracy on a 10-task subset of the DeepSWE software-engineering benchmark, against 65% for Claude Fable 5 and 52% for GPT-5.6-Sol over the same subset. Davis flagged the size of the sample himself: across ten tasks, one result moves the score by ten points. On the full 113-task benchmark, Ox Alpha came in around 63%, about level with GPT-5.6-Sol. No score for the model has been posted on any audited public leaderboard.
Whose model it was had already been worked out from the outside. Developers reached first for tokenizer fingerprinting, which examines how a model breaks text into the numeric pieces it operates on; researcher Joseph W. Elstner fed the model probe strings and, once he folded in the GLM-5 vocabulary Z.ai had published, all 95 matched with an average error of zero. Chetaslua then fired a malformed request at the model's server and got back an error identical to the one set out in Zhipu AI's own API documentation, rating the inference that Ox Alpha runs on Zhipu's infrastructure at 0.98 confidence.
The lab has run the play before, giving GLM-5 an unattributed preview on OpenRouter under the name Pony Alpha. It had put out GLM-5.3 on 14 August, six days before Ox Alpha turned up on the same platform with nothing on it to say where it came from.
A name on the provider does not settle what happens to the code already sent through it. The OpenRouter entry says the provider holds on to prompts and completions, with a pledge that they will not be used in training, and security researchers have flagged Ox Alpha as a third-party service that has not been vetted, urging developers to keep private code and secrets away from it.