On the morning of 20 August 2026, a new model appeared on OpenRouter with no lab name, no announcement, and no press release. The listing was called Ox Alpha, the developer field was anonymous, and the model card itself described it simply as a "stealth model." It had a 1M-token context window and, at launch, it was free in preview. Within 48 hours the spec had been copied into every AI news round-up of the week, along with a growing pile of theories about who actually built it.

This post separates what is in the record from what is rumor, and explains why the anonymity itself is the more interesting part of the story.

What is actually confirmed

The hard facts all come from the listing itself:

  • It exists and it works. Ox Alpha is a routable model on OpenRouter under a public endpoint. Anyone with an account can point a request at it and get generations back.
  • 1M-token context. The listing advertises a 1M-token context window — same class as the DeepSeek V4 Flash and Qwen 3.8-27B postings I have tracked here.
  • Free preview pricing. At launch the model was offered at $0 in preview, with community reports that the free window is expected to close around 27 August, after which a normal pay-per-token price would appear. I have not seen that date confirmed by anyone with standing to confirm it.
  • It is labeled "stealth." The model card's own language calls it a stealth model, and the developer identity is deliberately withheld. That is a choice, not an oversight — nobody accidentally publishes anonymously to a public API marketplace.

That is the entire confirmed list. Everything else is speculation, which is worth examining honestly.

The speculation

By 23–24 August, three theories had circulated:

  1. Z.ai's GLM. The strongest candidate in most commentary: Z.ai (the GLM lab) has a cadence of strong releases and a GLM successor has been expected. A 1M-context frontier-class model appearing anonymously fits the shape of an unannounced GLM launch. Note that my earlier GLM-5.2 coverage already put that lab's open-weights models at the top of agentic coding economics — a stealth release from them would not be a surprise.
  2. Microsoft's MAI. Less well supported. The theory rests on pattern-matching to Microsoft's in-house "MAI" line and on the fact that a 1M-context model is the kind of thing Microsoft has teased, but nothing in the listing points at Microsoft.
  3. xAI. The weakest of the three — it circulates mainly because an anonymous strong model always triggers xAI speculation, and I found no listing-specific evidence for it.

None of these is confirmed. The lab has not claimed the model, and no one has fingerprinted it publicly in a way I would trust before the free window closes.

Why a stealth launch matters

Stealth model launches are rare, and they exist for a specific reason. A lab that puts a strong model on a public router before a named launch is effectively running a blind eval on the open internet: real traffic, real workloads, no benchmark-gaming possible, and zero marketing spend. The anonymity protects the surprise while the eval runs. It also means the model is likely production-quality — stealth launches are how labs test that the serving stack survives real demand before the announcement.

For people building on APIs, this is also a concrete reminder: a strong anonymous model in your routing options can disappear or reprice with no notice. Treat it as a free trial, not a dependency. If you wire it into anything long-lived, keep a fallback model in the router.

The catch

To be straight about what I have and have not verified:

  • The identity theories are unconfirmed, and I am repeating them as theories, not facts. If a lab later denies one of the attributions, that section of this post was speculation from day one.
  • The "~27 August" free-window date is community-sourced, not from any listing I could attribute to the operator.
  • There are no independent benchmarks yet. With no named lab and no published eval suite, any performance claims floating around are anecdotal — and a 1M-context anonymous model at $0 will attract a lot of both flattering and adversarial testing over the next few days.
  • It may not survive. "Ox Alpha" is a listing name. When the launch happens, it will almost certainly get a real name, a model card, and real pricing — this post's facts will need a revision at that point.

What it means

The pattern this fits: the frontier labs are increasingly testing their next releases against the open market before they announce them, and the open-weights economics documented in my earlier posts on GLM-5.2 and DeepSeek V4 Flash mean an anonymous strong model at $0 is a legitimate probe of whether the pricing tier it is heading for will hold. The most useful thing you can do this week is run your own workloads at the endpoint and keep notes — once the free window closes, independent observations from the stealth period will be the only third-party record of how it actually behaved.

Sources

Related reading