Space Bunny Alpha: A Free 1M-Token Anonymous Model Nobody Has Claimed Yet

Space Bunny Alpha is an anonymous model that appeared today, September 23, 2026, on the third-party platform that hosts stealth previews: a 1,000,000-token context window, a 524,288-token maximum output, text, image and video input, tool calling, mandatory reasoning — and a price of zero. Nobody has said who built it. There are no published weights, no parameter count, no licence, no vendor page and no benchmark of any kind. This is not a release, and nothing in this piece should be read as one: it is a listing, and the useful work right now is reading the listing carefully.

That matters more than it sounds, because the last model to appear this way did not stay anonymous. Ox Alpha arrived on the same platform in August, was free for about a week, and turned out to be Z.ai's GLM-5.3-Flash — after which it was no longer free. Space Bunny Alpha is the newest entry in a sequence that has run all year, and the sequence is the most reliable thing we have to reason with while the identity is unknown.

What the listing actually says

Everything below is the listing's own metadata, read today. It is operator-supplied: the anonymous provider wrote it, and no third party has verified any of it.

• Context window — 1,000,000 tokens. Same 1M class as Ox Alpha's advertised 1,048,576.

• Maximum output — 524,288 tokens. Four times the 131,072-token ceiling Ox Alpha's August listing advertised, and unusually large for any model at any price.

• Input and output — text, image and video in; text out. Video input has become a signature of this series; most of the recent anonymous listings have taken it.

• Price — $0 per million tokens in and out. "This model is free to use," in the listing's words.

• Reasoning — mandatory, cannot be switched off, with a five-step effort ladder: low, medium, high, xhigh, max, defaulting to medium. A medium default is a change from the max-effort default the August listing used.

• Tool calling — tools and tool_choice are accepted. response_format gives JSON output, without JSON-schema enforcement, which is the same caveat as before.

• Data handling — prompts and completions "may be retained by the provider but are not used for training." Read that twice; it is a retention statement, not a zero-retention one.

• Undisclosed — parameter count, active parameters, architecture, tokenizer family (the listing says only "Other"), knowledge cutoff, quantisation, licence, and whether weights exist at all.

What the listing does not say

The absences are more informative than the numbers, because they are exactly where the previous models gave themselves away.

Ox Alpha leaked its own identity through a tokenizer that matched the GLM family probe-for-probe, a video encoder whose token behaviour was identical to a known Z.ai model, and a malformed request that returned an error trace naming the provider's own API route. None of that forensics exists for Space Bunny Alpha yet. Its tokenizer is filed as "Other", its knowledge cutoff is null, and no one has published a fingerprinting run — so any confident claim about which lab built it, including any you read this week, is a guess wearing evidence's clothes.

The second absence is measurement. There is no Artificial Analysis entry, no arena rating, no independent evaluation, and no community benchmark run that anyone has published. For a model with a 524,288-token output ceiling, that is the number a buyer would most want checked and the one nobody has checked: long-output coherence degrades well before a nominal ceiling, and a ceiling is not a capability.

The third is operational. The endpoint record publishes no throughput and no latency at all — both fields are empty.

The one claim the listing contradicts itself on

The description calls Space Bunny Alpha an anonymous large model "with blazing-fast inference". That is the operator's adjective. The same listing's endpoint record reports no output speed and no time to first token, and the tokenizer is undisclosed, so there is nothing on the page that supports the word and nothing that refutes it either.

It is worth naming because it is the single most common failure mode in anonymous listings: speed adjectives are cheap to write and expensive to verify, and the previous entry in this series was reported by early users as slow to first token and flaky under load even while it topped usage charts. Treat "blazing-fast" as a marketing string until somebody publishes tokens per second. If you are evaluating this model, measure it yourself on the first call — throughput and first-token latency are two numbers you can get in five minutes, and they will tell you more about whether it fits your workload than any identity guess.

The base rate: every one of these gets claimed

Since February 2026 the platform has run a steady sequence of anonymous Alpha listings, and every single one has ended with a named lab behind it:

• Pony Alpha (February) — Z.ai's GLM-5, claimed within about five days.

• Hunter Alpha and Healer Alpha (March 11) — Xiaomi's MiMo-V2-Pro and MiMo-V2-Omni.

• Elephant Alpha (April) — Ant Group's Lingxi Ling-2.6-flash, claimed roughly two weeks later.

• Owl Alpha (April) — Meituan's LongCat-2.0, confirmed in June.

• Ox Alpha (August 20) — Z.ai's GLM-5.3-Flash, revealed August 26, with the free window closing at the same moment.

• Union Alpha (September 16) — now listed alongside its named counterpart, Unbiased's Pareto.

Seven of seven, mostly Chinese labs, mostly within days to weeks. That is a base rate, not a prediction about Space Bunny Alpha — but it is the honest way to frame the guess: if you are reading this while the mask is still on, the most likely outcome is that a lab claims it soon, and the codename gives you no information about which one. The animal-and-Alpha naming convention has never correlated with the vendor.

The pattern carries one actionable detail. Ox Alpha's free window lasted about six days and closed on reveal. Free access in this series has been a preview instrument, not a permanent tier — so if you want to know what Space Bunny Alpha is good at on your own prompts, the cheap moment to find out is now, while the price is zero and the queue is short.

Trying an unproven model without betting production on it

Nothing about an anonymous listing stops you from evaluating it; what it should stop you doing is putting it on a path you cannot afford to have break. Two rules follow from the listing itself.

• Do not send it anything you would not hand to a stranger. The listing says prompts and completions may be retained by the provider. That rules out proprietary source, customer data and anything under a compliance regime. Public benchmark prompts, open-source code, and throwaway prototypes are fine.

• Do not make it the only model in a path. A provider with no name, no status page and no published latency is a provider whose failures you cannot escalate.

That second rule is what failover is for, and it is the part we can actually help with. OrcaRouter serves 200+ models behind a single API with no markup on provider list price — you pay the provider's own rate, so a vendor price cut is live on our side the same day — and automatic failover across providers. The routing DSL lets you declare a known-good model as the fallback for an experimental one, so an anonymous preview can absorb a real workload's traffic without being a single point of failure. Space Bunny Alpha is not on our catalogue today, so we are not a way to call this specific model; we are the way to keep a measured model behind it while you find out whether the anonymous one deserves the traffic. The last anonymous Alpha is a useful example of what the far end of that pipeline looks like: GLM-5.3-Flash is routable on OrcaRouter at its published rate, and its independent scores are public.

What would settle it

Three things, in rough order of likelihood.

First, tokenizer and error-path forensics. The community found Ox Alpha's identity in under a week with probes any developer can run; the same method applies here and has not been published yet. That is the fastest route from "anonymous" to "probably this family".

Second, an independent benchmark. A single full run on a public suite — not a ten-question smoke test, which is how the last anonymous model briefly acquired an 80% coding score that a full 113-task rerun knocked down to roughly 63% — would be the first real evidence about quality. Until that exists, the only quality claim in circulation is the operator's own description.

Third, the reveal itself, which in this series has arrived with a name, a licence, a parameter count and a price change all at once. When it comes, the questions that decide whether Space Bunny Alpha matters are unglamorous: are the weights open, what does it cost per million tokens after the free window, and does it hold quality across the output lengths its ceiling advertises.

For now the correct posture is the one the listing itself implies. A free 1,000,000-token window, 524,288 output tokens, video input and tool calling is a genuinely interesting thing to have on a key — and a genuinely bad thing to have on a critical path, from a provider that has told you nothing about itself except that it would like you to believe it is fast. Try it on your own prompts this week, keep a measured model behind it, and treat every identity claim you read until then as unverified. That is not caution for its own sake; it is the only reading the evidence supports.