Union Alpha: what we know about the unclaimed stealth model with a 256K window

Union Alpha is calling itself a frontier model. Nobody is calling it theirs. The name appeared as an anonymous stealth listing at 14:42 UTC today, September 16, 2026 — a 262,144-token context window, image input, tool calling, and a price of exactly nothing — with the provider field reading only "Stealth" and no lab, no model card and no weights attached to it. If the pattern of the last year holds, that silence resolves inside a week: Ox Alpha ran the same playbook on August 20 and was revealed six days later as Zhipu's GLM-5.3-Flash. This is a what-we-know-so-far piece. Everything specific to Union Alpha below comes from the listing's own metadata, published by an operator that has declined to name itself — the same metadata that describes "frontier-level performance across a broad range of general-purpose tasks," a claim the operator is making about itself and that nothing outside that listing reproduces.

So the honest version is short: a capable-looking model is free to call right now, it belongs to no one who will say so, and the only number published about it is one its own operator wrote. That is not nothing — stealth previews have a strong track record of turning into real, cheap, open-weight models — but it is a lead, not a result.

What the listing actually says, and what it scrupulously does not

The metadata is unusually specific for a stealth entry, which makes the gaps easier to read. Here is the whole of it:

• Context window — 262,144 tokens, with a maximum output of 131,072 tokens

• Input — text and images. Output is text only. There is no video input

• Price — $0 per million input tokens and $0 per million output tokens during the preview

• API surface — tools, tool_choice, response_format, temperature, top_p and max_tokens are all accepted

• Reasoning controls — none. No reasoning parameter is exposed at all

• Weights — no Hugging Face repository linked, no licence, no parameter count, no architecture note

• Knowledge cutoff — not stated. The field is empty

• Provenance — a single endpoint, provider name "Stealth", no failover provider behind it

Note what the spec sheet is missing rather than what it contains. There is no parameter count, so no way to tell a dense model from a mixture-of-experts, and no way to guess what it costs its operator to serve. The tokenizer is filed as "Other", which is the least useful possible value and forecloses the single most common piece of community forensics — the tokenizer fingerprinting that correctly identified Ox Alpha's lab in August. And the description's headline claim, "frontier-level performance", is operator-reported and unreproduced: no benchmark, no leaderboard entry, no third party has published a number for this model at the time of writing.

The absences are the interesting part

Read against Ox Alpha, the shape of Union Alpha is noticeably smaller — and not in a way that reads like a flagship held back.

Ox Alpha shipped a 1,048,576-token window and took video as well as images. Union Alpha offers a quarter of that context and stops at images. Ox Alpha exposed reasoning as a mandatory control with low, high and max effort levels, defaulting to max, and the thinking was impossible to switch off. Union Alpha exposes no reasoning control whatsoever — either the model does not reason in that explicit way, or the operator has chosen to hide the knob during preview. Ox Alpha also had a genuinely distinctive tell: it retained prompts and completions at the provider, a data-policy note that told you exactly what you were trading for free access.

A 256K window, text-and-image input and no reasoning tag is the profile of a fast tier or a sibling variant rather than a frontier flagship — the same relationship GLM-5.3-Flash turned out to have to the rest of its family. That is a reading, not a fact. It is worth stating plainly because the alternative reading is just as available: an operator trimming the advertised surface to keep a preview from being fingerprinted.

Every stealth model is somebody's, eventually

The stealth-preview format is now well-worn enough that the reveal is closer to a scheduled event than a surprise. The roll call, as reported rather than as confirmed by the labs themselves:

• Quasar Alpha and Optimus Alpha — turned out to be Ope​nAI GPT-4.1 previews in 2025

• Sonoma Sky and Dusk Alpha — x​AI's Grok 4 Fast

• Pony Alpha — Zhipu's GLM-5, February 2026

• Hunter Alpha — Xiaomi's MiMo-V2-Pro

• Owl Alpha — Meituan's LongCat-2.0

• Ox Alpha — Zhipu's GLM-5.3-Flash, revealed six days after it appeared

The Ox Alpha case is the one worth studying, because it shows both how well the crowd does and where it fails. Community testers called the lab correctly — tokenizer counts matched GLM-5.3 across a battery of prompts, the video encoder matched GLM-5V, and the API error formats were G​LM's. Several put their confidence near 99%. They still got the model wrong, assuming an unreleased multimodal variant of GLM-5.3 rather than a distinct "Flash" release, and the widely-quoted "80% on DeepSWE, beats GPT-5.6 Sol" figure that drove the first wave of coverage came from one developer passing eight of ten tasks on a subset — not the benchmark, and not comparable to the full-run scores it was printed beside.

When the reveal came on August 26, the real numbers were less dramatic and more useful: a 320B-parameter sparse mixture-of-experts with about 18B active per token, MIT-licensed weights on Hugging Face the same day, a $0.15 / $0.50 per-million-token list price with a 50% launch discount through September 9, and an Artificial Analysis Intelligence Index of 57 — an independent score, which is the thing a stealth preview never has while it is still stealth.

Who Union Alpha might be — and why the name is no help

The codename convention has been a weak but real signal. Pony, Ox, Owl, Hunter, Elephant: animal names, consistently, and the Chinese labs behind the recent wave have used them the way a studio uses a working title. "Union" breaks that pattern, and it does not map onto any known lab's naming scheme. It invites the reading that this is a joint or consortium release, which is exactly the kind of reading a codename is chosen to invite and which there is no evidence for.

No fingerprinting has been published for Union Alpha, and none is attempted here — one of the genuinely useful lessons from August is that confident forensics can name the right lab and the wrong model. What the first wave of testers will look for is predictable: token-count offsets against known tokenizers, error-string formats, refusal patterns, emoji rate, and whether a video encoder answers when one is fed in. Until something like that lands and survives a second look, any claim about who built Union Alpha is a guess wearing a confidence interval.

The part that should decide whether you touch it

Free access to an anonymous endpoint is not a bargain so much as an unpriced trade, and the terms are the thing to read carefully. You are sending prompts and completions to an operator with no name, which means no data-processing agreement, no stated retention period, no jurisdiction and no one to escalate to. There is no service-level agreement, no second provider to fail over to, and no notice period — a stealth endpoint can be withdrawn the moment its owner decides to speak, which is precisely what happened to Ox Alpha's free tier on reveal day. Anything built directly on that endpoint broke on somebody else's press release.

OrcaRouter does not route Union Alpha, and it would be wrong to imply otherwise — an anonymous endpoint with no named provider behind it is not something to put in a production path. What the routing layer is for is the step after. One API across the whole catalog — 198 models from 15 providers at the time of writing, at provider list price with zero markup — means that when a model like this one gets a real name and a real price — as GLM-5.3-Flash did on August 26 — it is a line item you can compare against everything else you already call, on the same key, the same day, without a second contract or a code change. Automatic failover and the routing DSL let you keep an unproven model behind a switch you control, so evaluating it never means betting a production path on it.

What to watch, and the one date that matters

If the recent pattern holds, the interesting window is the next five to seven days. Four things would turn Union Alpha from a listing into a fact: a name appearing in the provider field, weights landing in a repository under a stated licence, a published price, or an independent score on a leaderboard that does not take the operator's word for it. Each of those changes what you can do with it. A price tells you whether it is cheap or merely free. Weights tell you whether you can ever run it yourself. An independent benchmark tells you whether "frontier-level" was a description or a hope.

Two of those four would also make it routable, and the pass-through point matters there: when a vendor sets a list price we pass that price through with no markup, so a launch discount or a price cut is live on our side the same day it is announced rather than at the end of a billing cycle. That is the whole argument for keeping evaluation traffic on a router — you are not committing to a model, you are committing to a socket, and models can be swapped through it in a config line.

For now, the accurate answer to "how good is Union Alpha" is that nobody outside the anonymous operator knows. The window is real, the price is real, and the tool calling is real. The performance claim is a sentence the operator wrote about itself. That is worth a weekend of sandbox testing on code you would not mind losing, and it is not worth anything more than that until someone puts a name on it.