A model called Ox Alpha was released free on OpenRouter on Thursday, August 22, described as "a reasoning model designed for coding, sustained agentic work, and production workload". Its listing calls it a "stealth model" "developed and operated by a third-party provider who has chosen to remain anonymous during this preview", TechCrunch reports.
Patrick Collison, the chief executive of Stripe, which is acquiring OpenRouter, called it "very impressive".
Why a company would hide
Releasing a model anonymously is a recognized practice rather than a mystery. It lets a developer gather usage data and reactions without the release being read as a product launch, without a benchmark scoreboard forming around the company name, and without a disappointing showing costing anything. If it goes well, the name can be attached afterward.
That is the ordinary explanation. It is also the one that cannot be verified from outside, which is the whole point of the exercise.
The guessing
Attention first landed on GLM models from the Chinese company Z.ai. Some evidence was read as pointing to an unreleased version of Microsoft's MAI instead. On Reddit, one post argued it "can't be the Chinese" while another claimed "high confidence" that it is.
Andrew Curran, an AI analyst, noted that early speculation centered on Chinese origins but that "this morning people seem less sure of anything".
We are reporting that as a description of a conversation, not as evidence about the model's provenance. Two anonymous forum posts asserting opposite conclusions with equal confidence tell you about the state of the discussion and nothing about who wrote the model. We are not naming a likely developer, because there is no basis on which to name one.
The one hard number, and its limits
The article mentions that 100 trillion free tokens a day were being offered. It gives no performance benchmarks.
That figure, if it is a capacity offer rather than a measure of use, says something about the compute behind the model, and compute at that scale is not available to a hobbyist. It narrows the field to organizations with serious infrastructure. It does not identify one.
What we could not establish
We could not establish who built Ox Alpha, how it performs on any benchmark, whether the token figure describes capacity offered or usage recorded, or how long the preview will run. We verified this account from a single publication, and every attribution of origin in it is speculation by third parties.
