A new AI model showed up this week with a million-token context window, image and video input, and a name that sounds like a children’s cartoon. The company behind it? Secret.
That is the deal with Space Bunny Alpha, which appeared on OpenRouter and OpenCode on September 23. It is free during its preview, fast by OpenRouter’s reported measurements, and built for coding and reasoning. It briefly went offline while its provider fixed an issue; OpenRouter’s listing now shows it serving again.
Space Bunny is the newest entry in a growing AI ritual: release the model first, let developers test it under an alias, and reveal the maker later. The mystery brings attention. The free access brings traffic. And the developers supply the kind of feedback a polished launch demo cannot.
One model is a day old. The other is nearly a year old.
Space Bunny has been public since September 23, 2026. Its anonymous provider has disclosed little beyond its capabilities: a one-million-token context window, adjustable reasoning, and text, image, and video input.
Big Pickle, the other mystery model in OpenCode’s picker, has a much longer history. OpenCode’s model catalog dates it to October 17, 2025—more than 11 months ago. It has a listed 200,000-token context window and a 32,000-token output limit, and OpenCode still describes it as a free, limited-time stealth model. “Limited time” has been doing some heavy lifting.
Big Pickle also has a performance claim worth examining carefully. In August, an independent researcher reported that it resolved 63 of 124 tasks on Scale AI’s SWE Atlas Codebase QnA benchmark, or 50.8%. That figure exceeded the then-published GLM 5.2 result using the same Mini-SWE-Agent scaffold and the listed GPT-5.6 Sol result using Codex.
It is a researcher-run result, not an official Scale ranking. The run used one trial per task where Scale’s protocol uses three, and the GPT comparison used a different agent scaffold. It tells us Big Pickle was formidable on that August test. It cannot tell us with certainty which model was behind the alias then—or whether the alias serves the same model today.
That uncertainty is central to Big Pickle’s rumor mill. Developers have long suspected a Zhipu GLM-4.6 variant; others point to DeepSeek, particularly for later versions of the endpoint. OpenCode has not publicly pinned the current alias to either one. A mystery name that persists for 11 months may be a test bed, a changing route, or both.
The Space Bunny guesses just got more interesting
The first guesses included Moonshot’s next Kimi model. Some users tried the classic tactic of asking the model who made it. That produced conflicting clues, including references to OpenAI.
A more useful lead arrived September 24. An independent fingerprinting test sent 24 token-count probes through OpenCode’s Space Bunny route. Its counts matched the MiniMax models tested on the same gateway across all 24 prompts; the tested Kimi and GLM models matched less closely. That has pushed MiniMax, with a possible M3.1 or later M3-family preview, to the front of the speculation.
The distinction matters: a tokenizer match is evidence about a model family, not proof of its developer or exact version. Models can share tokenizers, and gateways add their own processing. The test was also run on OpenCode’s route, while OpenRouter lists Space Bunny under a separate model ID. MiniMax has not confirmed the theory.
This is what the online detective game looks like in practice. People probe knowledge cutoffs, compare coding behavior, inspect token counts, test prompts in different languages, and study how a model refuses requests. Asking “Who made you?” is entertaining, but a model can confidently invent an identity. A repeatable technical fingerprint is stronger evidence, even when it stops short of a reveal.
We have seen how badly first guesses can go. OpenRouter’s Pony Alpha appeared in February and was later identified as an early GLM-5 test. March’s Hunter Alpha drew DeepSeek speculation before Xiaomi confirmed it was an early MiMo-V2-Pro build. The crowd sometimes solves the puzzle. It sometimes gives the wrong company a free launch campaign.
Free tokens are part of the pitch
An anonymous preview lets a lab see how its model handles real coding sessions before its brand shapes users’ expectations. It can test demand and serving capacity while developers get to try a model without paying per token. Space Bunny’s brief outage shows the capacity test can be quite literal.
The bargain also depends on where you access the model. OpenCode says its Space Bunny provider follows a zero-retention policy and does not train on user data. OpenRouter says its Space Bunny provider may retain prompts and completions, although it does not use them for training. OpenCode separately says data collected during Big Pickle’s free period may be used to improve the model. Those are meaningful differences if an agent is reading your codebase.
This adds another dimension to the frontier AI price fight. Labs already compete on capability, speed, and the cost of a million tokens. Stealth previews put an attractive temporary price on the board: zero. For developers, the sensible question is less “Have I unmasked the bunny?” and more “Does it do my work well enough, and what happens to the work I send it?”
A reveal may eventually settle Space Bunny’s identity. Big Pickle suggests the more consequential question can stay open much longer: when you choose a model by a playful alias, how do you know what you will be using next month?
