Yesterday a new anonymous model showed up in OpenRouter’s catalog and we didn’t believe the price. Space Bunny Alpha —stealth/space-bunny-alpha — lists $0.00 input, $0.00 output. Not “free credits.” Free. And it isn’t a toy: 1,000,000-token context, up to 524,288 tokens of output in one reply, text + image + video input, and real tool calling with adjustable reasoning effort.
We read the spec straight off the catalog, wired it into both of our own agent rigs that same day, and started throwing real work at it. This post is what we found — and why the free tokens are the most underused resource in AI right now.
The Spec Sheet (Verified, Not Vibes)
Everything below is straight from the OpenRouter model catalog as of this morning. We don’t repeat vendor hype here; we read the machine.
A million tokens is roughly 750,000 words — your entire repository, a year of support transcripts, a few hundred pages of PDFs, dumped in a single request. Half a million output tokens means a model that will write you an entire documentation set, refactor a codebase section by section, or transcribe a long video into structured notes — in one go.
Its own one-liner in the catalog: “an anonymous large model with blazing-fast inference, strong coding capabilities and native multimodal input support.” That’s our read too: the reason it landed on our rigs as the default is that the cheap lane had become the useful lane.
What Free Tokens Are Actually For
The reflex is “free model equals disposable model.” Wrong. Paid-per-token APIs punish iteration, and iteration is how software gets built. This model inverts that: you can be reckless with experiments, because recklessness is now free. Concretely, it’s great at:
- Whole-repo review in one shot. Feed the codebase (no chunking, no RAG pipeline) and ask for a real architecture critique or a hunt for a specific bug class.
- Batch experiments that would be ruinous otherwise. A thousand-variant prompt sweep, a regression suite run through an LLM judge, a log-triage batch over months of files — at $0.
- Agent loops with tools.tools + tool_choice are supported, so you can point your own harness at it and let it run, with retries — the loop you’d hesitate to run on a bill.
- Multimodal triage. Drop in screenshots, UI flows, architecture diagrams or video frames and get one grounded answer instead of eyeballing hours of footage.
- Big-context grep days.“Where does this flag get set?” across an entire monorepo, in one request, for free.
Using It: One Curl, One Env Var
It speaks the OpenAI-compatible API through OpenRouter, so the wiring is boring — which is the point:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "stealth/space-bunny-alpha",
"messages": [{"role": "user","content":"Review this repo…"}],
"tools": [{"type":"function","function":{"name":"grep","description":"search files","parameters":{"type":"object","properties":{"q":{"type":"string"}},"required":["q"]}}}]
}'Pin the id in one place — MODEL=stealth/space-bunny-alpha in an env file — and point anything that speaks OpenAI at OpenRouter’s base URL. If you use one of the popular local agent CLIs, the same one-liner works there: the model id is the only thing that changes. (We set it as the default backbone in our own setup on day one, which is the fastest way to find out whether a model actually holds up.)
The Fine Print, Because It’s Free
Two honest caveats, stated plainly:
- Shared capacity. Free-tier endpoints are throttled and shared. Ship retries with backoff and a paid fallback, and never build a SLA on top of it.
- Stealth means stealth. It sits under the
stealth/namespace: per-model endpoints aren’t publicly listed (the direct model URL 404s), and availability can shift without a changelog. Treat it like a lab lane — great for building, not a promise.
The pro move is the hybrid: build and evaluate on the free lane, ship production on a pinned paid model — and use the savings on the paid side for the calls that actually need guarantees. You get the iteration speed for free and the uptime by choice.
Frequently Asked
Is Space Bunny Alpha really free?
As of September 24, 2026 the OpenRouter catalog lists $0.00 input and $0.00 output. It runs on shared free capacity, so rate limits apply and there’s no SLA. A lab lane, not a production promise.
How big is the context window?
1,000,000 tokens of context with up to 524,288 tokens of completion in a single response — about 750,000 words in, half a million out.
Can it look at images and video?
Yes — the modality is text + image + video in, text out. Screenshots, diagrams and video frames all go in with your prompt.
Does it support tool calling?
Yes — tools, tool_choice, reasoning_effort, response_format, temperature and top_p are all supported. Agent loops are the best way to spend free tokens.
Can I run production on it?
Don’t bet on it — prototype and batch on the free lane, pin production to a paid model, and always ship retries and a fallback.
Why does the direct model URL 404?
It’s a stealth/-namespace model: per-model endpoints aren’t public, and it only appears in the full catalog. Availability can change without a changelog — pin the id in an env var you control.
We Accept Donations
Send your crypto donations to Maxiaxxx.sol. All funds help towards moving our site to decentralized.