[Digital Today Seung-a Yoo, intern reporter] China's AI research lab Z.ai has been confirmed as the developer of the anonymous coding model Ox Alpha.

On Aug. 26 (local time), foreign media outlets including blockchain publication Cryptopolitan reported that Ox Alpha was uploaded to OpenRouter on Aug. 20 without a logo, company name or separate announcement. It drew millions of developer queries over a week.

Ox Alpha supports a context window of 1,048,576 tokens. It can process text, images and video and can output up to 131,072 tokens. That is a much wider range than typical models, which handle about 8,000 to 200,000 tokens.

The model was offered for free. Open-source terminal agent OpenCode introduced Ox Alpha the same day and said it would keep it free for a week, setting daily processing capacity at 100 trillion tokens. The OpenRouter page said an anonymous provider stores prompts and responses but does not use them for training. During this period, coding tools such as Claude Code sent corporate code to the model in the billions of tokens.

Early performance assessments drew strong interest. Developer Ben Davis (벤 데이비스) said Ox Alpha's first-try accuracy was 80 percent across 10 tasks on the DeepSWE benchmark. That was higher than Claude Fable5's 65 percent and GPT-5.6-Sol's 52 percent. He noted that the sample size was small, because on 10 tasks a one-task difference can move the score by 10 percentage points.

After expanding the evaluation to all 113 tasks, the score was tallied at about 63 percent. That is similar to GPT-5.6-Sol. No official leaderboard score from public verification has been released yet. Stripe CEO Patrick Collison (패트릭 콜리슨) also wrote on X, formerly Twitter, that the model was very impressive. Stripe agreed on Aug. 19 to acquire OpenRouter.

Developers and researchers traced the developer through tokenizer fingerprinting and server-response comparisons. Researcher Joseph W. Elstner (조지프 W. 엘스트너) said that after inputting a test string into Ox Alpha and then adding the GLM-5 vocabulary released by Z.ai, all 95 probes matched with an average error of 0. Analysis using Modelprint, created by developer unclecode, also found that 6 of 9 infrastructure probes matched GLM-5.3.

Researcher Chetaslua (체타슬루아) confirmed that the error format returned after sending a corrupted request matched Z.ai's official API documentation. A test using an invalid role value also produced the same error as Z.ai's GLM model. When the same weights were served on a different host, the error format differed, and he rated the confidence of the estimate at 0.98.

Z.ai previously used the name Zhipu AI. GLM-5 was also previously released anonymously on OpenRouter under the name Pony Alpha, and GLM-5.3 was launched on Aug. 14.