Ox Alpha: Free model processes 23.2 trillion tokens on OpenRouter

Starting on August 20, developers could try an anonymous AI model for free on OpenRouter. Ox Alpha explicitly asked for feedback and within a few days rose to the top of OpenRouter's weekly ranking by processed token volume: Through August 25, it received around 295 million requests, while approximately 23.2 trillion tokens were processed in total.

The appeal was not just the price. OpenRouter lists a context window of around one million tokens — enough room for large amounts of source code and previous work. Ox Alpha can also process images and videos and call external tools. For AI agents that write software while interacting with other programs, this lowers the barrier for lengthy, complex tests.

The initial hype was driven by a small comparison: Ox Alpha solved eight of ten tasks in the DeepSWE coding benchmark, giving it a score of 80 percent. The fully documented run across all 113 tasks, however, produced a score of just 58.4 percent. That is within the range of established commercial models, but below the top scores of Claude Opus 5 at 74 percent and GPT-5.6 Sol at 73 percent. The community runs are also only partially comparable because of differences in test configurations.

In concrete terms: Developers were able to test a capable model with a long context and tool access on real programming tasks without an access barrier. For Z.ai, the covert launch likely also provided insights into the model's strengths, weaknesses and typical use cases. An independent assessment by Artificial Analysis is still pending; the figures available so far are therefore a signal, not a final verdict.

The model's origins have now been clarified. The open-source tool Modelprint found technical similarities with GLM-5.3, including in the tokenizer; Z.ai confirmed to Bloomberg that Ox Alpha is a new variant of its GLM series. This confirmation also makes the privacy issue concrete: Inputs and responses were sent to a Chinese company. Z.ai said it would publish the model weights on Wednesday and offer hosted access free of charge for one week in total; a later price had not yet been determined.