OpenRouter posted today that a new stealth model called Union Alpha, from an account called @unionalphaai, is live on its platform: free to use, a 256K (262,144-token) context window, tool calling, image input, and — in the model's own words, relayed by OpenRouter — "frontier-level general-purpose performance." No lab is named. Nothing about @unionalphaai identifies a company behind it.
What's actually verified here
Worth separating cleanly, the way any stealth-model post on this blog should: the context window, the price, and the tool-calling support are listing metadata OpenRouter itself publishes and can be checked directly against the API. "Frontier-level general-purpose performance" is a sentence the model's own operator wrote about itself, with no benchmark attached, no comparison set, and no third party behind it. That's not a small distinction — it's the entire distinction this blog has applied to every self-graded capability chart from a named lab this year, and an anonymous operator's unsupported adjective deserves at least the same skepticism as a named one's chart.
Why this one is harder to fingerprint than the last one
This blog covered the last major stealth-model mystery in detail: Ox Alpha appeared on OpenRouter and OpenCode on August 20, and within two days two people, using two unrelated technical methods, had independently converged on the same answer. One matched Ox Alpha's token counts against several candidate models across 15 probe strings and found an exact match with Zhipu's GLM-5.3 tokenizer, off by a constant 75-token wrapper offset, while every other candidate model missed badly. The other triggered a malformed-request error that leaked a server-side class name matching Zhipu's own documented API route. Z.ai confirmed the identity directly six days later.
Union Alpha closes off the first of those two paths before anyone can try it: OpenRouter's own listing files its tokenizer simply as "Other" — not a named, comparable tokenizer family, which is the actual raw material tokenizer fingerprinting needs to work. No published attempt at fingerprinting Union Alpha has surfaced as of this writing, and without a named tokenizer to compare against candidate models' known vocabularies, the method that resolved Ox Alpha's identity in under 48 hours doesn't have anywhere to start. Whether that's a deliberate choice by whoever built Union Alpha, aware of exactly how Ox Alpha got traced, or simply an artifact of a genuinely different underlying architecture, isn't something this post can settle from the outside.
The meta-lesson from Ox Alpha, worth restating precisely
It's worth being exact about what actually resolved the last mystery, because the record gets rounded up in retelling. The tokenizer and stack-trace fingerprinting pointed at Zhipu convincingly, days before anything else did — but what actually confirmed it was Z.ai's own statement: "Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback." Community fingerprinting produced a well-evidenced leading hypothesis; a first-party disclosure is what turned it into a fact. That pattern reportedly holds across every prior stealth-model reveal this cycle has produced, not just Ox Alpha's: fingerprinting has repeatedly pointed at the right lab in advance, and has never once been the thing that actually settled the question on its own. Worth keeping that distinction sharp for Union Alpha specifically, given there isn't even a fingerprinting lead to weigh yet.
A pattern that's accelerating, not a one-off
Ox Alpha in August wasn't the first anonymous stealth release this cycle, and Union Alpha isn't the last one already circulating — a separate stealth model called Omen Alpha reportedly appeared on OpenCode around the same time this month, pricing itself at $0.20 per million tokens rather than free. Community tooling has caught up to the trend: an open-source project called modelprint now exists specifically to automate the kind of infrastructure fingerprinting that took individual sleuths a day or two to improvise for Ox Alpha by hand. The stealth-release-then-community-guessing-game cycle has gone from a novelty to routine enough that there's now off-the-shelf tooling built for it — worth noting as its own small data point about how normalized this launch pattern has become in under a month.
Three things are worth watching rather than assuming. Whether any published fingerprinting attempt surfaces at all — a "tokenizer: Other" listing closes off the fastest method that worked last time, not necessarily every method. Whether any hypothesis that does accumulate gets treated as a leading guess rather than a resolution, since Ox Alpha's own identity was only actually settled by Z.ai's own words, not by the fingerprinting that preceded them. And whether a disclosed, reproducible benchmark ever attaches itself to "frontier-level general-purpose performance" — the same standard this blog has held every named lab's launch chart to this year applies here with less information to work from, not more license to relax it.