HunterAlphaHub
OpenRouter model reference Facts from the public catalogue, dated and labelled
Jev Laya Fable 5.2 Verified
2026-11-13

Stealth line · delisted, unclaimed

What Was Space Bunny Alpha? The Anonymous Model That Topped OpenRouter

Space Bunny Alpha appeared on OpenRouter on 23 September 2026 with no announcement, no maker and no price — the fourth anonymous model in the Alpha line, after Hunter Alpha (later Xiaomi MiMo-V2.5), OX Alpha (later Z.ai GLM 5.3 Flash) and Union Alpha (later Unbiased Pareto). It ran for 13 days, spent most of them as the busiest model on OpenRouter by token count, and left on 2026-10-05 — the date its own listing had been printing for a week. Unlike those three, nobody has said who it was.

So this page is no longer a tracker; it is the record of a codename that got away. The status route below still answers honestly, and it will go on answering: fetch it and it says the id does not resolve.

Record Ran 2026-09-23 → 2026-10-05 Verified 2026-10-06 Delisted 2026-10-05, no claim since OpenRouter listing →
Checking OpenRouter… Reading the public catalogue.

The short version

Maker
Never named
strongest lead: MiniMax, unconfirmed
Listed
23 Sep 2026
14:48 UTC
Delisted
2026-10-05
13 days after it appeared
Usage
#1 on OpenRouter
28.4T tokens, its best week
Context
1,000,000
tokens
Max output
524,288
tokens
Input
Text + image + video
output is text
Price while up
Free
$0 in / $0 out, every day of it
Claimed by
Nobody
no maker has said a word
What it became
Unknown
the question a reveal would answer

How much it was being used

On 2026-10-01, a week into a thirteen-day run, OpenRouter's own rankings page listed this model first on Top LLMs by weekly token usage on OpenRouter — ahead of vendors with names. The figures below are the platform's, not ours, and they are the last ones it published before the id went away.

# Model Tokens this week
1 Space Bunny Alpha 28.4T
2 DeepSeek V4.1 Flash 22.7T
3 GLM 5.3 Flash 10.6T
4 MiMo-V2.6-Flash 9.1T
Requests, 2026-09-30
505,254,305
the platform's own count for one day
Prompt tokens that day
27.4T
input side only
Completion tokens that day
978.6B
output side

The leaderboard counts tokens, and a model that costs nothing is used heavily for that reason alone. Read it as attention, not as a verdict on the model.

Verified specs

Read straight from the public catalogue and its endpoint payload — both re-checkable in one request. Rows that other people published rather than we confirmed are marked as such, and there are none on this page today.

FieldValue
Model id
stealth/space-bunny-alpha
no longer resolves
Listed
2026-09-23 (14:48 UTC)
the catalogue's own `created` field
Delisted
2026-10-05
the listing's banner said “Going away October 5, 2026”; gone by 2026-10-06, 03:00 Beijing
Context window
1,000,000 tokens
1M
Max output
524,288 tokens
the endpoint's cap
Input modalities
text, image, video
output: text only
Price while anonymous
$0 in / $0 out
per million tokens, for the whole run
Provider
Stealth
OpenRouter's placeholder for an unnamed vendor
Provider disclosure
None
no maker named, before or after the delisting
Source openrouter.ai/api/v1/models Verified 2026-10-06 Live status /api/space-bunny-alpha/status

What the listing declared

Four things stood out while it was up, and all four were checkable. None of them is a performance claim.

1M-token context, 524K output cap
The window is the largest on this site — the same 1M as Hunter Alpha and OX Alpha — and the output cap is four times the 128K the last codename allowed.
Video as an input modality
The catalogue declares text, image and video in, text out. Union Alpha took text and images; this is the first codename in the line that also names video.
reasoning_effort is exposed
Alongside `reasoning` and `include_reasoning`, so the caller can ask for less or more deliberation per request. The model card lists it as a feature rather than a default.
Tools and response_format are both supported
So the route can be asked for structure — the thing every prior codename also offered, and the reason these models get dropped into agent loops before anyone benchmarks them.

The terms, in one sentence

Prompts and completions for this model may be retained by the provider but are not used for training; all other use is governed by the Stealth Model Terms.

The provider is anonymous *and* may keep what you send. That is not the same as training on it, and it is not nothing: if the state you send contains personal data, this is the sentence that decides whether you may. The listing and its terms →

What the delisting did not answer

FieldValue
Who made it
Not public, and the delisting did not disclose it. OpenRouter's endpoint payload named the provider only as “Stealth”. The MiniMax evidence below is the strongest public lead and it is not a claim by anyone involved.
Whether it will ever be named
Unknown, and this line has no precedent for silence: every earlier codename was claimed by its maker within days of the delisting, or before it. This is the first one to leave without a word.
What a reveal would look like
Either OpenRouter's own page for the codename gaining a “revealed to be …” sentence, as Union Alpha's did, or the maker's model appearing in the catalogue under its real name — usually a new id created in the days around the delisting.
How it performed
We ran nothing against it and publish no latency, cost or accuracy figure of our own. The listing's own adjectives are not a benchmark, and it is too late to benchmark it now.

The identity lead

No maker has claimed this model — not while it ran, and not in the days since it left. The listing called the provider nothing but "Stealth" for its whole life. What exists instead is evidence, and it points one way: MiniMax. The lead is below with the confidence it deserves and the reason it could still be wrong. It is not in the spec table above, because the listing never stated it.

MiniMax — most plausibly a preview of MiniMax M3.1 Flash Medium confidence

Two independent counts published 2026-09-24, run through OpenCode's route for this model and for named vendors: a 24-pair test matched MiniMax in every pair and matched Kimi and GLM far less well; a 50-string test matched MiniMax on all 50. Four days later, MiniMax added M3.1-Flash-Preview to MiniMax Code with the same 1M-token window, the same three input modalities and reasoning that cannot be switched off. None of that is proof — token counts show family resemblance, a serving gateway can rewrite a token stream, and MiniMax has not claimed the model. Why the theory exists at all: It is the only lead with evidence rather than a hunch, and the timing fits: MiniMax put a same-shaped, unnamed model into its own tool four days after this one appeared. Two editors reading the same evidence would reach the same name.

A different lab, on a MiniMax-compatible tokenizer Low confidence

Token counts show technical kinship, not ownership. As of 2026-10-01 OpenRouter still routes this under the provider name "Stealth", MiniMax's own site does not mention the codename, and no lab has claimed it. Why the theory exists at all: Every identity guess has a way to be wrong, and this one's is the tokenizer itself: tokenizers get shared. It stays on the page so the MiniMax line reads as the best available answer rather than the answer.

Timeline

  1. 2026-09-23T14:48:04Z
    Listed on OpenRouter
    `stealth/space-bunny-alpha` appears in the catalogue with no announcement and no maker. The timestamp is the catalogue's own `created` field; public posts date it to the 23rd or 24th by timezone.
  2. 2026-09-24
    The first identity test
    A group of researchers runs 24 paired token counts through OpenCode's route, for this model and for known vendors. It matches MiniMax closely in every pair and Kimi and GLM far less well. A second outfit, YFarmX, runs 50 strings the same day and reports 50 of 50. Token counts show kinship, not ownership.
  3. 2026-09-24 → 2026-09-27
    Write-ups, still no maker
    A Chinese-language news item, two aggregator pages and a post from OpenRouter itself. None names a vendor, and no repository or build we can find uses the model yet.
  4. 2026-09-27
    MiniMax slips out an unnamed sibling
    `M3.1-Flash-Preview` appears in the model picker of MiniMax's own Code tool with no announcement, no price and no benchmarks — the same 1M-token window, the same text+image+video input, and reasoning that cannot be switched off. It is the closest public match the codename has, and it is not a confirmation.
  5. 2026-09-27
    Our first read of the listing
    1M context, 524,288 max output, text+image+video input, $0 in and out, `reasoning_effort` among the supported parameters.
  6. 2026-09-28
    Still live, and the press reports no claim
    Still listed, still free, 99.97% endpoint uptime over the previous 30 minutes. Chinese-language business coverage follows the identity hunt and states plainly that MiniMax has not claimed the model.
  7. 2026-09-30
    It becomes the most-used model on the platform
    OpenRouter's usage feed records 505,254,305 requests and about 28 trillion tokens in a single day, and its weekly leaderboard puts the codename first, ahead of DeepSeek V4.1 Flash and GLM 5.3 Flash.
  8. 2026-10-01
    The listing prints an end date
    The model page gains a banner the earlier readings did not carry — "Going away October 5, 2026" — the first hard expiry in this line. Still anonymous, still free, still unclaimed by any maker.
  9. 2026-10-06
    Delisted, and still unclaimed
    The id stops resolving. A watcher that has been reading the catalogue every three hours since 1 October puts the change between 00:00 and 03:00 Beijing time on the 6th — the 5th in the vendor's timezone, the day its own banner named. OpenRouter now lists no `stealth/*` id at all, and its page for this codename still says "anonymous": there is no "revealed to be" line as there was for Union Alpha. Nobody has said whose it was.

How to check all of this yourself

Read /api/space-bunny-alpha/status — it re-reads the catalogue on every request and answers whether the id resolves, whether it is still free, and what the endpoint declares. Since 5 October it answers that the id does not resolve, and that is the claim worth checking rather than taking from us. Or fetch openrouter.ai/api/v1/models and search it for stealth/space-bunny-alpha: you will not find it, and the fields on this page are the fields that payload carried while it was there.

Supported parameters, as it declared them: reasoning, reasoning_effort, include_reasoning, max_tokens, temperature, top_p, tools, tool_choice, response_format.

FAQ

What is Space Bunny Alpha?

It was an anonymous model on OpenRouter, listed as stealth/space-bunny-alpha from 23 September 2026 to 5 October 2026. The catalogue described it as a large model with fast inference, strong coding and native multimodal input, a 1M-token context window and adjustable reasoning effort, and on the week it peaked it was the busiest model on the platform by token usage. It is gone, and nobody has claimed it.

Is Space Bunny Alpha still available?

No. The id stealth/space-bunny-alpha stopped resolving on 5 October 2026 (confirmed by a catalogue read in the small hours of the 6th, Beijing time), and OpenRouter currently lists no stealth/* id at all. It was free for its whole run — $0 in and $0 out, the normal state of a stealth preview — and it left without ever being repriced, because it left without being claimed.

Was it revealed?

Not yet. Every previous codename in this line was claimed by its maker around the delisting, and OpenRouter's page for those codenames carries a sentence saying what they turned out to be. This one's page still calls it anonymous, no new MiniMax model has appeared in the catalogue, and no lab has said a word. The delisting was the deadline the listing itself printed; the reveal is a separate event that has not happened.

Can I send private data to it?

Read the terms first. The listing says prompts and completions may be retained by the provider and are not used for training. Retained is not the same as trained on, and it is the sentence that decides whether the state you send may leave your systems.

Who makes it?

Still unnamed, and now it may stay that way. OpenRouter's endpoint payload called the provider “Stealth”, its placeholder for a vendor that has not been disclosed. The strongest public lead is MiniMax — two independent tokenizer counts published on 24 September matched MiniMax where they matched nothing else, and MiniMax added a same-shaped M3.1-Flash-Preview to its own tool on 27 September. That is a lead with evidence, not a confirmation, and the page keeps it in the claims section rather than in the spec table.

How long did it last?

Thirteen days: listed 23 September 2026, delisted 5 October 2026. It is the only codename in this line that printed its own end date — Union Alpha ran two days and vanished without one, Hunter Alpha ran months. The status route still answers honestly rather than being retired with the model: /api/space-bunny-alpha/status reports whether the id resolves right now, and since the 6th it has said it does not.

Why was Space Bunny Alpha at the top of the rankings?

Because it was free, and good enough that people kept using it. OpenRouter's own weekly leaderboard, read 2026-10-01, put it first by token usage at 28.4 trillion tokens for that week, ahead of DeepSeek V4.1 Flash at 22.7T and GLM 5.3 Flash at 10.6T; the same feed recorded 505 million requests for it on 30 September alone. That measures attention, not quality — a model that costs nothing will always rank above one that does not — but it is real evidence that the preview was worth running.

Where this fits

The stealth index lists every codename we can tie to a listing, with what each one turned out to be. The Alpha Line Report is the register of the 3 that have been revealed — Space Bunny Alpha is deliberately absent from it until it has an after, because a before-and-after table with half a row is a different table. What stealth models are explains the pattern itself.

And the OpenCode page is the correction this one needed: "delisted" is true of OpenRouter and false of OpenCode, which still serves this codename free — and sells it by the token on its Go plan.

Get notified if this one is ever named — or when the next codename appears

This one has already done the part that usually comes with a reveal — the codename is gone — without the reveal. If the maker ever comes forward, or if the next codename shows up in the catalogue, one email says so, with what we could verify.

One email per reveal. No newsletter, no forwarding, unsubscribe in one click.

Send us a link

A project built with Jev, a post, a video, a correction, a tip. We open the link, check it says what you said it says, and write the entry ourselves.

Required a link, and an email to reply to. Optional everything else.

Add context — all optional

One or two sentences about what it does, in your words. We write the entry ourselves.

Cost, latency, a benchmark — anything you measured. We attribute these to you.

We store what you type and email it to ourselves. No IP address, no user agent, no referrer — the same rule as the mailing list.

What happens to what you send →