OpenRouter Stealth Models: Who They Turned Out to Be

OpenRouter stealth models are anonymous, usually free listings that a lab runs for a week or two under a codename before the real model ships. Fourteen of them have appeared since April 2025. This page is a census of all fourteen: what each one was called, when it listed, what it turned out to be, who said so, and how strong the evidence behind that identity actually is.

The census exists because the question that matters about a stealth listing is not “what is it” but “will we ever find out.” The answer is a base rate, and the base rate can only be read off a complete, dated list. That list did not exist in one place: the rows are scattered across OpenRouter’s blog, its retired model pages, two labs’ own announcement channels, and a handful of press reports, and the secondary compilations that do circulate are missing rows and carry stale identity statuses.

What follows is the method, the full table, a chart of how the fourteen slots resolved, the six historical rows without an official reveal, the data terms each listing carried, the dated evidence on the one row still open, and instructions for redoing the work. The launch-day story of that open row, Ox Alpha, is told in our August 20 post and is not repeated here.

A figure has circulated this week, drawn from a single secondary compilation, that OpenRouter had run eleven stealth slots before Ox Alpha and that three of those eleven were never revealed; the twelve-row framing counts Ox Alpha itself as the twelfth row. Going back to OpenRouter’s own blog, X account and model pages for every row, and to the underlying labs’ own channels where a reveal was claimed, changes that count in three places.

The corrected recount, across the thirteen historical rows: seven officially revealed, four with a community identity theory that no vendor has ever stood behind, and two, Cypher Alpha and Aurora Alpha, with no credible identity theory at all. Ox Alpha is the fourteenth row and is open. The base rate for “we will find out” is materially better than the circulating framing suggested, and the reason is instructive: three of the missing reveals came from the underlying lab rather than from OpenRouter, and a compilation that only watches OpenRouter’s blog will miss exactly that class of event.

That also qualifies the pattern most coverage repeats, that reveals come from OpenRouter’s own blog. The accurate version is that a reveal, when it comes, always arrives through an official first-party channel: OpenRouter’s blog or X account for four rows, and the lab’s own blog or X account for three. Fingerprinting has never yet been the mechanism of a reveal.

A census is only citable if a stranger can redo it and land on the same numbers. The block below states the collection method, the sources, the as-of date, the count, the exclusions and the known limits, and separates cells that were read off a primary from cells that rest on someone else’s inference.

Three evidence tiers are used throughout. VERIFIED means a first-party statement names the model: OpenRouter’s own reveal post, or the lab’s own announcement. REPORTED means the listing facts are verified from the model page but the identity rests on community inference with no vendor statement found. Unresolved means no identity theory has credible evidence behind it, and Open is reserved for a listing still running on the as-of date.

The table is the asset. Listing names link to the model’s own OpenRouter page where one was read; reveal mechanisms link to the first-party statement or the best available corroboration. Listed dates with a slash (2025-04-02/03) are carried as the primary records give them, across two days. Context is the figure on the model page or API record, in tokens. Data as of August 22, 2026.

Two cells deserve a note on what they do and do not claim. Row 8’s tier reads “Reported (weak)” because no vendor statement of any kind exists for Polaris Alpha; the GPT-5.1 identity rests on leak-based speculation alone, which is thinner than the OpenRouter model record behind the two Sonoma rows. Row 14’s “turned out to be” cell deliberately says nothing about the leading theory; a census cell is not the place for an identity no vendor has confirmed, and the evidence is laid out with dates in section 07 instead.

The chart below is the evidence-tier column of the table, counted. It is the single most citable figure on this page because it is the direct, dated answer to the question every stealth listing raises: how often do we actually find out.

Where a reveal did come, it came inside a fortnight in six of seven cases. Optimus Alpha was named 4 days after listing, Horizon Beta after 6, Hunter and Healer Alpha after 7, Horizon Alpha after 8, and Quasar Alpha after 11 to 12, depending on which side of the day boundary its listing is counted from. The outlier is Owl Alpha: listed 2026-04-28 and confirmed 2026-06-30, a 63-day gap, and the one reveal that a watcher monitoring only OpenRouter’s own channels would have missed entirely, because it came from Meituan.

The honest reading of the chart is two-sided. A slot that resolves usually resolves inside a fortnight and always through a first-party channel; but six of the thirteen historical slots never resolved at all on the evidence available, and two of those never acquired even a credible theory. “We will find out soon” is a better bet than the circulating framing implied, and it is still not a safe assumption to build a delivery date on.

These are the rows where the “will we find out” thesis fails, and they stay in the table for that reason. Each card states what the community believes, what the primary record supports, and what was searched for and not found. A dedicated search pass was run for every one of the four REPORTED rows; no xAI or OpenAI statement turned up for any of them. Six historical rows are covered below, across five cards, because the two Sonoma listings were a same-day pair; Ox Alpha is added as a sixth card because it is open on the as-of date rather than unresolved.

A pattern worth recording: the four REPORTED rows all have the same shape. Two presumed xAI builds in September, one presumed OpenAI snapshot and one presumed xAI reasoning build in November, and in every case the presumption rests on routing behaviour, output style or leaks rather than on anything either lab has said. Neither lab has confirmed or denied, and absent a statement these rows will stay REPORTED indefinitely. That is not a defect of the census; it is the finding.

Every historical listing’s data-terms notice was read off its own OpenRouter page. The wording varies cosmetically; the scope does not. The secondary table below groups the fourteen rows by exact wording so the variance is visible cell by cell. Data as of August 22, 2026.

The general policy those thirteen defaults sit under is the Stealth Program terms at openrouter.ai/terms/stealth, marked “Last Updated 2026-07-06” on the page. It grants OpenRouter a licence, described as irrevocable and perpetual, to sublicense user content to stealth providers “for the sole purpose of enabling the Stealth Provider(s) to train, evaluate, and improve those Stealth Model(s).” The privacy backstop is a hashed identifier: “User Content provided to Stealth Providers will contain a hashed identifier, such that each individual user will not be identified or identifiable to the Stealth Provider,” with a separate contractual bar on providers attempting re-identification. The terms also state plainly that there is no selective opt-out: not wanting content used this way means not using stealth models at all.

Ox Alpha is the exception to that policy, and OpenRouter’s own announcement on X framed it as one rather than as a change: “this time, the provider does not train on your prompts or completions.” Three qualifications travel with that sentence. Retention still occurs; the model page says prompts and completions are retained, just not trained on. The promise is a policy statement, not a technical guarantee anyone outside can verify. And the API record sets is_moderated: false, so no moderation layer sits in front of the model. How these notices compare with the terms coding agents publish is the subject of our sibling census of agent data terms.

stealth/ox-alpha was created in OpenRouter’s catalog at 2026-08-20 20:04:55 UTC. The API record states a 1,048,576-token context, 131,072 maximum completion tokens, text, image and video input, $0 prompt and completion pricing, mandatory reasoning with three effort rungs (max, high, low; no medium) defaulting to max, and a null knowledge-cutoff field. It is the only model in the 422-model catalog whose id or name contains “stealth”; every other Alpha-named slot in the table above has been retired or renamed out of the live API, although their documentation pages remain live.

Everything else known about the row is community evidence, and the table below dates each item. Two framings are excluded on purpose. The “100 trillion tokens per day” figure is not an OpenRouter number and not a provider number; it is OpenCode’s marketing copy about capacity on its own resale route, and it is recorded here only with that attribution. And the model has no audited benchmark: the DeepSWE figures are two community subset runs, reported together, with the correction that accompanied the second.

One consequence is worth projecting forward. If the row does resolve to a Z.ai model, anyone who built on the free slug during the preview inherits a pinning problem: the slug retires, the named model lands at list price, and the ~z-ai/glm-latest floating alias will repoint to whatever Z.ai ships next. The census can tell you the reveal is likely; it cannot tell you the date, and the precedent for the free endpoint afterwards is retirement.

To reproduce the table: pull https://openrouter.ai/api/v1/models and filter for “stealth” in id; separately browse https://openrouter.ai/blog/announcements/ for “stealth,” “cloaked” and “Alpha” titles; for each historical candidate name, fetch https://openrouter.ai/openrouter/<slug> directly, since most retired stealth models keep a live documentation page after the model itself stops serving; and cross-reference every identity claim against the named lab’s own blog or X account before elevating a row from REPORTED to VERIFIED. A row moves tiers on a first-party statement and on nothing else. For what a listing date does and does not tell you about a model, see our catalog-literacy guide; for what OpenRouter lists in an ordinary month, the July catalog roundup is the baseline stealth slots are an exception to.

This page is maintained in place under a stable slug. When Ox Alpha is revealed, row 14 changes tier and the chart is recounted; when a new stealth slot lists, a row 15 is added. The as-of date in the methodology block and the table caption moves with each refresh, and the modified date in the page metadata records it. If you re-fetch a primary later and a cell has changed, the change is the news; cite the as-of date alongside the figure. Teams that route production traffic through catalogs like this one, and need the listing-versus-launch distinction built into their model selection, are the kind of work covered by our AI and digital transformation practice.