China's Z.ai emerges as creator of mysterious Ox Alpha model

Chinese lab Z.ai has been identified as the company behind Ox Alpha, the anonymous coding model. 

The model drew millions of developer queries on OpenRouter over the past week. 

Who created Ox Alpha? 

The Chinese lab called Z.ai has been credited with the creation of Ox Alpha, a coding model that has exploded in popularity over the past week, drawing millions of developers.

The model was released on OpenRouter on August 20, carrying no logo, no press release, and no company.

The Ox Alpha listing boasted a massive context window of over a million tokens (1,048,576) that could handle text, images, and video, and spit out up to 131,072 tokens. Most models can only handle 8,000 to 200,000 tokens.

Access to the model costs nothing. OpenCode, an open-source terminal agent, launched the model the same day and said it would stay free for a week, advertising provider capacity of 100 trillion tokens per day.

OpenRouter’s page says the mystery provider keeps prompts and completions but promises not to use them for training. Despite being completely anonymous for over a week, coding tools like Claude Code pushed billions of tokens of company code through it.

Ox Alpha went viral when the first run a developer called Ben Davis completed put the model at 80% first-pass accuracy on a 10-task slice of the DeepSWE software-engineering benchmark, ahead of Claude Fable 5 at 65% and GPT-5.6-Sol at 52%.

Davis flagged the small sample himself. On a 10-task set, one task swings the score by 10 points. When he finished the full 113-task benchmark, Ox Alpha landed near 63%, roughly on the same level with GPT-5.6-Sol. However, no audited public leaderboard has posted a score.

Stripe CEO Patrick Collison, whose company agreed on August 19 to acquire OpenRouter, tried the model and called it “very impressive” on X.

How did researchers find out who created the model?

The first tool developers used to try to identify the source was tokenizer fingerprinting, which reads how a model splits text into the numeric units it processes.

Researcher Joseph W. Elstner ran probe strings through Ox Alpha and, after adding Z.ai’s released GLM-5 vocabulary, matched all 95 of his probes with a mean error of zero.

A separate developer, unclecode, built a fingerprinting tool called modelprint and found Ox Alpha lined up with GLM-5.3 on six of nine infrastructure probes.

A researcher named Chetaslua confirmed the model was running on Zhipu’s system after he sent a broken request to its server and received an error message that matches Zhipu AI’s official API documentation.

A bad-role test also returned the same error that Z.ai’s GLM models produce. When the same weights were served by a different host, the error format changed, confirming the suspicion. Chetaslua rated the inference at 0.98 confidence.

Z.ai, which was formerly Zhipu AI, also previewed its GLM-5 on OpenRouter anonymously, using the name Pony Alpha. GLM-5.3 shipped on August 14, six days before Ox Alpha appeared.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.