Ox Alpha’s token count has moved from unusual to industrial-scale. In an August 24 post, OpenRouter said the anonymous model was on track to process nearly 6 trillion tokens that day. The platform’s latest completed leaderboard, updated through August 24, shows 17.5 trillion tokens over the previous week.
The numbers align with the model’s rapid ascent. Business Insider had reported 11.6 trillion tokens during Ox Alpha’s first two days. Adding another 5.9 trillion brings that earlier snapshot close to OpenRouter’s current 17.5 trillion total.
Its maker remains unconfirmed. Community fingerprints lean toward Z.ai and the GLM model family, but nobody has produced conclusive evidence, and no AI lab has publicly claimed Ox Alpha.
Ox Alpha Is on Track for a Six-Trillion-Token Day
OpenRouter’s “today” referred to Monday, August 24, rather than Tuesday, August 25. The completed leaderboard now gives that projection more weight: the difference between the earlier 11.6 trillion figure and the newer 17.5 trillion total is approximately 5.9 trillion tokens.
Spread evenly over 24 hours, 6 trillion tokens works out to roughly 69 million tokens per second across all concurrent requests. It does not mean one instance of Ox Alpha was generating at that speed. OpenRouter aggregates traffic from many applications, developers, coding agents, and automated processes using the model simultaneously.
Ox Alpha’s lead is still extraordinary within that context. Its 17.5 trillion tokens are more than 13 times the 1.33 trillion recorded by second-place MiniMax M2.5. The anonymous model also processed about three times as many tokens as the other nine models in OpenRouter’s top 10 combined.
Raw token volume can rise quickly in agentic software development. A coding agent may repeatedly submit repository files, conversation history, terminal output, test results, and tool responses. One programming task can therefore produce dozens of calls and consume far more tokens than a conventional chatbot conversation.
Free Access Helps Explain the Usage Explosion
The Ox Alpha model card describes it as a text model optimized for software engineering and agentic coding. It has a 131,072-token context window, accepts text input, produces text output, and is currently offered without an input or output token charge. OpenRouter lists its creation date as August 21, 2026.
That combination is built to attract heavy experimentation. Developers can connect a coding harness, give the agent a large codebase, and allow it to run multi-step tasks without watching a metered API bill increase after every attempt. OpenRouter’s post specifically promoted using Ox Alpha through its ori command-line tool with existing coding agents.
Free access also makes the leaderboard harder to interpret as a market comparison. The 17.5 trillion figure measures consumption, not revenue, customer retention, benchmark performance, or willingness to pay. A free anonymous model can draw curiosity traffic, automated evaluation runs, and unusually long agent sessions that would be expensive on a normal commercial API.
Developers should also examine the privacy terms before submitting proprietary code. OpenRouter’s current provider information says prompts and completions may be retained, although they are not used for model training. Free access should not be confused with a guarantee that sensitive repository data is never logged.
The GLM Fingerprints Are Suggestive, Not Conclusive
Ox Alpha arrived without an organization name, technical report, benchmark table, license, or detailed architecture description. Its provider label is simply “Stealth,” leaving users to infer its origin from output style, tool behavior, coding performance, and occasional model self-identification.
Much of the speculation points toward Z.ai, formerly known as Zhipu AI, and its GLM family. Community testers have reported similarities in response patterns and agentic behavior. Some outputs have also appeared to identify the underlying model as a GLM variant, although asking an LLM to name itself is not a reliable attribution method. Model outputs can reproduce names found in training data, evaluations, or system-level instructions.
A separate line of speculation focuses on political prompts. Participants in a Reddit testing discussion reported that Ox Alpha would discuss topics such as Tiananmen Square and criticism of Xi Jinping. Some users interpreted that behavior as evidence against a Chinese developer.
That conclusion goes too far. A Chinese-trained foundation model offered through an international evaluation endpoint could use a different system prompt, safety layer, or deployment policy from a domestic consumer chatbot. Its willingness to answer sensitive questions may reveal something about this specific deployment, but it does not establish where the base model was trained.
The strongest responsible assessment is therefore narrow: behavioral fingerprints lean toward GLM, while the model’s actual developer remains unknown.
“Niu Lai” Is a Meme, Not a Lab Attribution
Chinese social media users have given Ox Alpha a more colorful identity while the technical investigation continues. Business Insider traced the nickname “Niu Lai” to the animated movie Here Comes the Ox, whose Chinese title uses the same ox-related wordplay.
The film follows a mother-and-son pair of yaks and became an online meme after attracting attention for its unusual production history and animation. It subsequently earned about 47 million yuan, or approximately $6.7 million, turning a widely mocked release into an unlikely box-office curiosity.
Because Ox Alpha appeared without an official identity, users connected “Ox” in the model name with the viral movie. “Niu Lai” stuck, while others began calling the model the “Bull Demon King.”
The nickname is useful internet folklore, not forensic evidence. Anonymous models almost always attract theories based on codenames, logos, response quirks, and cultural references. Those clues can guide testing, but a pun cannot prove that Z.ai, another Chinese lab, or any particular company created the model.
Six Trillion Tokens Do Not Prove Model Superiority
Ox Alpha’s traffic shows that developers will eagerly test a capable-looking coding model when the financial barrier disappears. It also shows how much LLM usage agentic programming can generate once a model starts operating inside automated loops rather than answering isolated chat prompts.
The leaderboard does not tell us how many people used Ox Alpha, how the traffic was distributed, or whether a small number of high-volume applications produced most of the tokens. It does not provide independent coding benchmarks, failure rates, latency statistics, or comparisons conducted under controlled conditions. OpenRouter’s ranking covers its own platform rather than the entire AI market.
There is also no announced open-weight release. “Free on OpenRouter” describes the current API price, while “open weight” would mean downloadable parameters accompanied by a license defining how they can be used. Ox Alpha may eventually be revealed as a preview of an open-weight GLM model, but the available evidence does not establish that.
The more revealing test will begin after the stealth period. If the model receives normal pricing, will developers continue sending it trillions of tokens? If its identity is disclosed, will independent evaluations support the positive coding reports? Those answers will separate durable demand from the temporary effects of novelty and free access.
Final Thoughts
Ox Alpha’s identity mystery is entertaining, particularly now that Chinese users have attached it to a viral cow movie. The usage figures are more consequential. Moving from 11.6 trillion tokens after two days to 17.5 trillion through August 24 amounts to one of the clearest demonstrations yet of how quickly free coding agents can generate infrastructure-scale LLM demand.
The reveal alone will not settle the important questions. Ox Alpha still needs reproducible benchmarks, transparent pricing, clearer data-handling information, and an official statement about its developer and licensing. Until then, 6 trillion tokens in a day is evidence of intense experimentation, not proof that the anonymous model is the best coder on OpenRouter.
Frequently Asked Questions
5 questions
1What is Ox Alpha on OpenRouter?
Ox Alpha is an anonymous LLM offered through OpenRouter’s Stealth provider. It is designed primarily for software engineering and agentic coding, supports a 131,072-token context window, and is currently free to use. OpenRouter has not identified the company or research lab responsible for developing the model.
