Space Bunny Alpha: Inside OpenRouter's Free Anonymous AI Experiment

A new name slipped quietly onto the AI scene on September 23: Space Bunny Alpha. The anonymous model appeared out of nowhere on the OpenRouter and OpenCode listings, free to use for the whole week. In the video its name keeps coming out as SpaceMoney, but the real listing reads Space Bunny Alpha, and that name confusion sits at the very center of the story told here.

How the stealth preview works

The most striking thing about the model is not what it does but how it is served. Developer and operator is a third-party provider that chose to stay anonymous for the preview. OpenRouter merely routes requests to it and is not its developer, owner, or provider. Under the terms, prompts and completions may be retained by the provider but are never used for training. OpenRouter's Stealth page formalizes the arrangement: Union Alpha arrived September 16 with 262K context, Ox Alpha on August 20 with 1.05M. The precedent matters because Ox Alpha was later revealed as GLM-5.3-Flash, which suggests today's mystery may be solved one day too; SioraLabs documented that reveal.

One million tokens and mandatory reasoning

The spec sheet reads surprisingly generous for a free preview. A 1M-token context window accepts text plus native image and video input, returning text. Adjustable reasoning effort runs across five levels, max, xhigh, high, medium and low, defaulting to medium. The ceiling is 524,288 completion tokens with tool calling on top, so the model is shaped for agentic workflows rather than chat. SioraLabs' live API scan confirms zero-dollar pricing on input and output with full uptime. The host adds his own measurements to the table: 74 tokens per second and 1.5 seconds to first response. Those two are unverified, so read them as claims.

Tokenizer fingerprints and the collapsed MiniMax theory

The identity hunt began fifteen minutes after launch. The first clue was a tokenizer fingerprint : one account reported the text tokenizer matching MiniMax M3 exactly. MarMar Labs ran the careful version, 36 input comparisons across 12 fixtures, and every one matched the public MiniMax M3 tokenizer, with a self-declared catch that M2.7 matched too, pointing at the family rather than a release. YFarmX counted 50 test strings and found all 50 splitting exactly like MiniMax. CellCog's September 27 deep dive holds the line: the MiniMax traces are strong but the identity is unconfirmed.

The theory wobbled when MiniMax officially launched M3.1-Flash-Preview on September 27. According to PANews the model went live on MiniMax Code, aimed at everyday development, fast bug fixes and stable responses, sweetened by a check-in campaign paying double points from September 28 to October 7. The host shelves the theory and declares the hunt back to zero. CellCog stays cooler: the new release shares the stealth model's published specs, MiniMax remains plausible, but nobody has named it.

The voxel garden duel: 45 minutes and zero dollars

The duel stage is a detailed voxel Japanese garden built with 3JS: a small temple, tiny villagers, a flying dragon and interactive touches. Both models got the same prompt. Space Bunny Alpha finished in about 45 minutes at top reasoning effort and the run cost nothing. GPT-6 Astra in X-High thinking mode finished in about 15 minutes but the generation cost $4.24. Against his own expectations the host calls the free model's output more detailed and more careful with small touches. The rival's portrait is worth recording: OpenAI credits GPT-6 Astra with 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench and a 59.3% lead on the Agents' Last Exam. Duel numbers are the host's measurements; the OpenAI scores come from the official announcement.

A web showcase tour: from headphones to an OS

The heavy lifting ran inside VS Code through the Codium coding agent with the model pinned to maximum reasoning. Task one was a 360-degree product viewer: the headphone model came out genuinely three-dimensional, clickable parts opened feature cards, color switching worked. The 3D threshold where most models stumble was cleared. Task two was a macOS-style operating system running in the browser: glassmorphism interface, notification center, display and sound settings, night mode, top-bar menus and SVG-icon apps all generated. Clicking around left flicker bugs across the windows, but the skeleton worked. The third showcase was an Nvidia graphics-card landing page: after four attempts a spinning-fan 3D card, SVG logo, typography and dynamic motion came together.

The price of speed and honest failures

Not every experiment made the reel. A Call of Duty-style request landed 2D sprites inside a 3.js scene; the enemies looked odd and no real 3D clone emerged. The globe task drew a harsher table: Space Bunny Alpha took 43 minutes and 219K tokens while GPT-6 Sol answered in about 5 minutes on 520K tokens. A name correction belongs here too: the model called Soul in the video is actually Sol from OpenAI's Sol and Luna announcement. Top reasoning effort buys quality but inflates wall-clock time; a model that feels fast in normal modes stretches minutes toward hours at max.

The model that watches video: the multimodal peak

The strongest demo came where the model used its eyes. Handed a game clip and asked to rebuild it in HTML, it reproduced the storm effect, pen-sketch ships, the Black Pharaoh boss read straight from the game UI, the collection system, bombs, the win screen and WebAudio-generated sounds. At maximum effort the first token arrived after about 7.8 seconds and 50K tokens streamed at roughly 120 per second. Original clip on top, rebuild running below, in the same frame. A similar reading feat turned a floor-plan image into a full 3D apartment through a Blender link in about an hour. The host adds that the hidden checkpoint keeps getting faster and better than its first days, apparently updated on incoming feedback.

The closing round asked for a complete stock-research app from a single prompt: market data fetching, charts, competitor mapping, SEC filings and recent events gathered into one clean responsive interface with bull, base and bear scenarios. The claim is that the model made most decisions itself, from data source to which rivals to include. Access is straightforward: the model runs through OpenRouter and inside OpenCode with whichever coding agent you prefer. OpenCode's docs describe 75-plus providers, model picking via the /models command, and recommended picks including GPT 5.2, Claude Opus 4.5 and Minimax M2.1. A 3D-packaged landing page came out in a single sitting as well. Trying it before the free window closes makes sense; watching what you paste into it is mandatory.