stealth/space-bunny-alpha has just doubled its share of OpenRouter’s image category: 5.3% to 11.2%, up 5.8 points, for the week of 2026-09-28. openai/gpt-5.6-sol entered the same ranking at 4.1%.
Nobody official will say whose model it is. It has no card, no lab and no paper — and it is reported to be withdrawn today.
First, what is actually known — and what isn’t
OpenRouter runs a “stealth” category. Models appear under a stealth/ namespace, free or near-free, operated by a provider that has chosen to remain anonymous during a preview period.
On space-bunny-alpha specifically, third-party write-ups report: listed 23 September 2026, a 1,000,000-token context window, maximum output of 524,288 tokens, fast inference, strong coding, native multimodal input — and a withdrawal date of 5 October 2026, which is today.
Identity: tokenizer tests reportedly point to MiniMax, with specs matching an M3.1-Flash-Preview. This is inference by third parties, not confirmed by OpenRouter or any lab.
And one thing is unresolvable from the available sources: OpenRouter’s own page for the model publishes no modality, context length, pricing, provider or sunset date. The ranking it appears in is OpenRouter’s image category, but the third-party descriptions emphasise multimodal input rather than image generation — so whether 11.2% represents image generation share or share of traffic involving images cannot be established. The share is real; its denominator is not documented.
That uncertainty is not a footnote. It is the condition the whole category operates in.
What the stealth arrangement is for
It is a reasonable deal for the provider and it is worth being clear-eyed rather than cynical about it.
You get real-world evaluation at scale, for free. Benchmarks do not tell you how a model behaves against the long tail of what people actually ask. A week on OpenRouter against thousands of real users and real prompts does.
You get it without the reputational exposure of a launch. If the model is bad, nothing was announced and nothing has to be retracted. If it is good, you launch with evidence.
You avoid comparison on your competitors’ terms. No name means no leaderboard entry, no head-to-head coverage, no narrative about whether you have caught up.
And the pricing is honest about what is happening. Free access in exchange for usage data is the oldest arrangement on the internet. Users understand it.
What it costs the people who use it
Here is where it matters for anyone building things.
You cannot cite it. If a piece of work, a research result or a product decision rests on a model with no name, no version and no provider, it is not reproducible by anyone — including you, next month.
You cannot depend on it, and the withdrawal date proves the point. A model that takes 11% of a category’s traffic and then disappears on a Monday leaves every workflow built on it broken. Anyone who spent the last fortnight tuning prompts against space-bunny-alpha’s particular behaviour has today lost that work.
You cannot assess the terms. No published provider means no published data-retention policy, no jurisdiction, no statement about whether your prompts are used for training. For personal experiments that is a shrug. For a client’s material, a collaborator’s unreleased work, or anything with a person’s likeness in it, it is not.
And the free price distorts the measurement. An 11.2% share for a free model does not mean 11.2% of users preferred it. It means it was free and fast. Share under those conditions measures availability, not quality — which is precisely why the provider wants the number and why you should not read it as a verdict.
The practical position
Use stealth models for exactly what they are: a free look. They are a genuinely good way to find out whether a capability is close, and to form an opinion before the marketing arrives.
Do not put one in anything that has to keep working. If a stealth model is the best tool for your job, treat that as a signal to watch for its named release, not as a dependency to build on.
And write down what you are using, always. A note of the model string, the date and the observed behaviour costs nothing, and it is the only thing that will let you make sense of your own results after the model is gone.
The broader read: this is now a routine part of how frontier models reach the public, and the version of the story that matters is not which lab space-bunny-alpha belonged to. It is that a significant share of real usage in a category can run through a model whose provider, terms and lifespan are all undisclosed — and that the field has come to treat that as unremarkable.