An unidentified AI lab has released a mysterious reasoning model called Ox Alpha, offering users access to up to 100 trillion tokens per day for an entire week at no cost. The unusually high limits have triggered speculation across the AI community about the identity of the company behind the model.
Ox Alpha reportedly appeared anonymously on OpenRouter and OpenCode on August 20, featuring a 1-million-token multimodal context window, a 131,072-token output limit, and support for text, image, audio and video inputs.
Ox Alpha's Capabilities Fuel AI Community Speculation
According to its own description, Ox Alpha is designed for coding, sustained agentic tasks, production workloads, long-horizon software engineering, complex reasoning and workflows involving visual information.
The model's claimed infrastructure capacity is particularly notable. The anonymous provider says it can process as much as 100 trillion tokens per day, with the offer remaining free for seven days. Such a scale has led observers to question which AI laboratory could have the computing resources required to support the deployment.
Evidence Points Toward China's Zhipu
Among the leading theories, Zhipu has emerged as a potential developer of Ox Alpha. One reason for the speculation is Zhipu's previous practice of testing unreleased models under alternative "Alpha" names. Its GLM-5 model was reportedly live-tested under the Pony Alpha moniker.
Several technical similarities have also been cited by observers:
- Video encoding pipelines and tokenizer calculations associated with Ox Alpha reportedly resemble Zhipu's architecture.
- Ox Alpha has allegedly produced character-for-character responses similar to those generated by Zhipu models.
- On third-party benchmark platform Kingbench, Ox Alpha reportedly scored 87.5%, placing it below the 91.25% score attributed to GLM-5.3.
Ox Alpha Could Be an Unreleased GLM-Class Model
The speculation becomes more significant because GLM-5.3's publicly available endpoints reportedly do not offer native vision or video capabilities, while Ox Alpha supports multimodal inputs, including text, images and video.
If the Zhipu theory proves accurate, Ox Alpha could therefore represent an unreleased and substantially upgraded GLM-family model, potentially offering capabilities beyond those currently available through Zhipu's public models.
However, the identity of Ox Alpha's developer has not been officially confirmed, leaving the model's origins—and whether it is actually connected to Zhipu, DeepSeek, xAI or another AI laboratory—unresolved.
.png)
