OpenRouter put a new anonymous model in its catalog at 14:48 UTC on September 23, 2026 and called it Space Bunny Alpha. It is free, it takes text, images and video, it holds a million tokens of context, and nobody has claimed it. Four days later it is one of the most talked-about coding models on X, and the only named company whose product matches it is MiniMax.
This page is the record: what OpenRouter published, what the fingerprinting found, how Space Bunny lines up with the M3.1-Flash-Preview model MiniMax put in its coding app on September 27, and what nobody has confirmed. A disclosure first: we build CellCog, an AI employee platform, and we do not route to Space Bunny Alpha or any stealth endpoint.
On this page · 10 sectionsOpen
OpenRouter listed Space Bunny Alpha as stealth/space-bunny-alpha at 14:48 UTC on September 23, 2026: a 1,000,000-token context, up to 524,288 output tokens, text, image and video input, reasoning that cannot be switched off, and a price of $0 in and out.
Nobody has claimed it. OpenRouter says a third-party provider built and runs it and chose to stay anonymous for the preview; OpenRouter only routes the requests.
Three independent tokenizer tests posted on launch day point to MiniMax: one account called it M3.1 outright, one matched 36 of 36 comparisons to the MiniMax M3 tokenizer but stopped at the family, and one matched 50 of 50 strings.
On September 27 MiniMax posted that M3.1-Flash-Preview debuts on MiniMax Code. Its config lists the same three traits as Space Bunny: a 1,000,000-token context, thinking forced on, and effort levels from low to max.
It is fast in OpenRouter’s own numbers: 79 tokens a second and 1.61 seconds latency at the median, 98.96 percent availability over three days. No vendor benchmark exists, because there is no named vendor.
The free window has a clock. OpenCode announced it as free for a week on September 23. Every earlier OpenRouter stealth model lost its free price the moment its maker was named, and Union Alpha lasted 33 hours.
§ 01What OpenRouter published
The listing describes it as “an anonymous large model with blazing-fast inference, strong coding capabilities and native multimodal input support.” On who is behind it, the page is explicit: “It is developed and operated by a third-party provider who has chosen to remain anonymous during this preview.” OpenRouter’s launch post called it “A flash model with fast inference, adjustable reasoning, and a 1M-token context window.” Everything below comes from OpenRouter’s public model API and listing page, read at 14:01 UTC on September 27.
| Item | Value |
|---|---|
| Model id | stealth/space-bunny-alpha |
| Listed | September 23, 2026, 14:48:04 UTC (catalog created timestamp) |
| Maker | Anonymous third-party provider |
| Context and output | 1,000,000 tokens in, up to 524,288 out |
| Input | Text, image and video; text out |
| Reasoning | Mandatory; efforts low, medium, high, xhigh and max, default max |
| Tools | Tool calling (auto only) and structured output |
| Price | $0 input, $0 output |
| Speed, week median | 79 tokens a second, 1.61 seconds latency |
| Tool call error rate | 3.16 percent |
| Availability, three days | 98.96 percent |
§ 02The MiniMax clues, graded
The identity work started within 15 minutes of the launch post. At 15:03 UTC, the account cheaty wrote: “New stealth model ‘Space Bunny’ is most certainly MiniMax M3.1.” Its test compared token counts: “I can independently confirm that the text tokenizer perfectly matches that of the MiniMax M3.” Twenty-six minutes later it added: “Vision tokenizer is a perfect match too, this is most certainly MiniMax-M3.1!”
The more careful read came from MarMar Labs at 16:32 UTC: “My current bet on Space Bunny Alpha is MiniMax. M3.1 is a plausible candidate, but I can’t call the exact model yet.” It made 36 input-token comparisons across 12 text fixtures and every one matched MiniMax’s public M3 tokenizer, with a catch it named itself: “M2.7 matched too, so this points to the family rather than a specific release.” The Qwen, GLM, DeepSeek and Llama tokenizers it tried did not match. YFarmX reached the same family at 17:52 UTC: “We counted how its tokenizer splits 50 test strings. All 50 match MiniMax exactly.”
On September 26 the benchmark account LuminaBench went further, posting that a MiniMax deployment it calls M3-a returns the name space-bunny in its benchmark runs, and predicted: “This comes out of stealth testing Wednesday, i’d expect it to drop the same day”. We cannot check that deployment from outside.
| Claim | Source | Grade |
|---|---|---|
| Made by MiniMax | Three tokenizer tests, September 23 | Consistent evidence, unconfirmed by MiniMax or OpenRouter |
| Specifically M3.1 | cheaty; MarMar Labs calls it plausible | Plausible, specs match M3.1-Flash-Preview, unconfirmed |
| A MiniMax deployment returns space-bunny | LuminaBench, September 26 | Single source, cannot check |
| Leaves stealth on Wednesday, September 30 | LuminaBench | Prediction |
| Free for a week | OpenCode, September 23 | Stands; the week ends around September 30 |
| Not trained on your prompts | OpenRouter listing and post | Stands; retention still allowed |
§ 03Space Bunny next to M3.1-Flash-Preview
At 05:24 UTC on September 27, MiniMax’s agent account posted: “MiniMax’s latest text model, M3.1-Flash-Preview, debuts today on MiniMax Code.” It added: “Built for everyday development, it’s fast, reliable, and ready for real work, from quick bug fixes to full features.” MiniMax Code’s public model config lists the model, and its three published traits line up with the OpenRouter listing one for one. The older M3, in the same config, does not.
| Trait | Space Bunny Alpha (OpenRouter) | M3.1-Flash-Preview (MiniMax Code config) | MiniMax M3 (same config) |
|---|---|---|---|
| Context | 1,000,000 tokens | 1,000,000 tokens | 450,000 tokens |
| Reasoning | Mandatory | Forced on | Switchable, on by default |
| Effort levels | Low, medium, high, xhigh, max | Low, medium, high, xhigh, max, plus default | Not listed |
| API price | $0 during the preview | Not published | $0.30 in, $1.20 out per million tokens on OpenRouter |
A match on three traits is a strong hint, not a name. M3.1-Flash-Preview lives only inside MiniMax’s coding app for now: MiniMax’s model release notes still end at July 31, its Hugging Face organization has nothing newer than August 7, and OpenRouter carries no M3.1 id. MiniMax has published no benchmark for it.
§ 04How the week went
Demand ran ahead of any explanation. OpenCode’s post announcing the free week had passed 990,000 views by September 27, and the posts under the model’s name are mostly builders showing apps made from a screenshot or a sketch.
§ 05The stealth playbook, and its record
OpenRouter has named every earlier stealth model, and the free price ended each time.
| Model | Appeared | Revealed as | Days anonymous |
|---|---|---|---|
| Quasar Alpha | April 3, 2025 | An early GPT-4.1 test | 11 |
| Horizon Alpha | July 30, 2025 | An early GPT-5 checkpoint | 8 |
| Pony Alpha | February 6, 2026 | Z.ai GLM-5 | 5 |
| Ox Alpha | August 20, 2026 | Z.ai GLM-5.3-Flash | 6 |
| Union Alpha | September 16, 2026 | Pareto, by Unbiased | 1.4 |
| Space Bunny Alpha | September 23, 2026 | Open | 4.0, counting |
The last two runs are the useful comparison. Ox Alpha was fingerprinted as GLM from its tokenizer and the fingerprint was right on every count when Z.ai named it six days in. Union Alpha fooled the fingerprinters, because it was a blended model from a startup nobody had guessed. Space Bunny looks like the Ox Alpha case: a clean tokenizer match to one lab’s family, and now a same-spec product from that lab.
§ 06The data question
OpenRouter’s launch notes say: “This time, the provider does not train on your prompts or completions”. The listing is narrower: “Prompts and completions may be retained by the provider but are not used for training”, with everything else under OpenRouter’s Stealth Model Terms. No training is a real promise; it is not a promise of deletion, and the company holding the data is not named. OpenCode describes its own route as keeping no data, which is a claim about its route, not about the OpenRouter endpoint. For code or documents you would not hand to an unknown company, wait for the name.
§ 07What we are watching for
- A name. A statement from OpenRouter, MiniMax or another lab, or the stealth listing losing its endpoints, which is how Union Alpha’s reveal showed up in the catalog.
- An M3.1 API listing. A MiniMax model id on its platform, on OpenRouter or on Hugging Face, with a price, and whether its specs match the stealth listing.
- Numbers. A MiniMax model card for M3.1-Flash-Preview, or an independent coding run on the stealth endpoint.
- The end of the free week. OpenCode’s week runs to about September 30, the day LuminaBench expects the reveal.
- The stealth-week prompts. What the named maker says about prompts retained during the anonymous window.
§ 08The tracker
| Question | Status |
|---|---|
| Who made it | Unnamed; evidence points to MiniMax |
| Which model | M3.1-Flash-Preview matches three published traits; unconfirmed |
| Is it free | Yes, $0 in and out on OpenRouter |
| Benchmarks | None from a vendor; OpenRouter speed and error rates only |
| Data terms | Retained, not used for training, per the listing |
| Reveal | Open |
§ 09Update log
- September 27, 2026, 14:10 UTC. Page opened, four days after the listing and nine hours after MiniMax’s M3.1-Flash-Preview post. Listing facts from OpenRouter’s API and page; posts read on X with times checked against their ids; MiniMax Code config read from its public endpoint.
§ 10Sources
- OpenRouter: Space Bunny Alpha listing
- OpenRouter launch post, September 23
- cheaty, tokenizer match, September 23
- MarMar Labs, 36 comparisons, September 23
- MiniMax Agent: M3.1-Flash-Preview on MiniMax Code, September 27
- MiniMax Code model config
Q1Is Space Bunny Alpha the same as MiniMax M3.1-Flash-Preview?
Probably, but it is not confirmed. The tokenizer evidence points to the MiniMax family, and the two share a 1,000,000-token context, thinking that cannot be switched off and effort levels from low to max. MiniMax has not said the OpenRouter endpoint is its model, and M3.1-Flash-Preview has no API listing, weights, price or benchmarks yet.
Q2How do people identify a stealth model?
By tokenizer fingerprinting: they send test strings and compare the input-token counts the endpoint reports with the public tokenizers of known model families. A match rules families in or out, but models in one family often share a tokenizer, so it rarely names the exact release. That is why one tester called it MiniMax M3.1 and another stopped at the MiniMax family.
Q3What happens to prompts I send to Space Bunny Alpha?
OpenRouter’s listing says prompts and completions may be retained by the anonymous provider but are not used for training, with everything else under OpenRouter’s Stealth Model Terms. OpenCode described its own route as zero data retention. Treat anything you send as readable by a company you cannot name yet.
Q4Is Space Bunny Alpha good?
Early users praise it for fast coding and multimodal input, and OpenRouter measures 79 tokens a second at the median with a 3.16 percent tool call error rate. There is no vendor benchmark and no independent run we can verify, so quality claims are anecdotes until a named release publishes numbers.
Q5When will Space Bunny Alpha be revealed?
Nobody official has said. One benchmark account predicted it leaves stealth on Wednesday, September 30, the day OpenCode’s free week would end. Past OpenRouter stealth models were named after 1.4 to 11 days.