An anonymous model called Space Bunny Alpha has spent the past several days sitting atop the daily usage charts on both OpenRouter and the open-source coding tool OpenCode. OpenRouter labels it only as coming from a third-party provider that "chose to remain anonymous during the preview period" — the platform routes requests but is not the model's developer. As of Sept 28, no vendor has stepped forward to claim it.
What's publicly known about the specs
According to third-party compilations of OpenRouter's listing records, Space Bunny Alpha went live on Sept 23 at 14:48 UTC — 10:48 p.m. Beijing time that night. A handful of specs have been confirmed:
- A 1 million token context window, with a per-response output cap of 524,288 tokens;
- Input supports text, images and video; output is text only;
- Supports tool calling and JSON-formatted output;
- Both input and output pricing are zero.
OpenCode announced free access for one week on launch day. Runtime data on OpenRouter's page shows a median output speed of about 79 tokens per second, a first-token latency of roughly 1.61 seconds, three-day availability of 98.96%, and a tool-call error rate of 3.16%.
QbitAI ran a round of coding tests: a Mac clock app took 22 minutes and 46 seconds to build, a screen-caption tool took under 5 minutes, and the cache-hit rate stayed above 95% on all but one task. Testers described it as fast, usually solving problems in a single pass, though its reasoning traces run long and it tends to build extras beyond what was asked for.
The clues point to MiniMax
Online speculation about the model's identity has at various points touched OpenAI, xAI, Meta and several domestic vendors. The most concentrated set of clues right now points to MiniMax.
One developer ran a tokenizer fingerprint comparison on launch day, and all three independent tests matched MiniMax's M3 series. On Sept 27, MiniMax rolled out M3.1-Flash-Preview on its own coding product, MiniMax Code. The documentation describes a 1 million token context window, support for text, image and video input, reasoning that can't be turned off, and a five-tier effort parameter for adjusting reasoning depth — all of which overlap heavily with Space Bunny's observed behavior, and both are positioned as speed-oriented Flash-tier models.
All of this remains circumstantial. Neither OpenRouter nor MiniMax has confirmed the two are the same model, and M3.1-Flash-Preview still has no published API pricing or benchmark scores, so outsiders currently have no way to directly match the test data on both sides.
What domestic developers can take from this
Launching under an anonymous codename on OpenRouter to rack up free usage before revealing its identity is a path domestic vendors have already tried. Before Meituan's LongCat 2.0 went open source, it was quietly offered for free under the codename Owl Alpha, and its usage briefly topped the platform. For a vendor, it's a public stress test free of brand expectations — real usage data arrives faster than internal evaluations do.
For users, the free window comes with two caveats worth weighing. First, the time limit: OpenCode states the free period lasts only one week, and going by past anonymous-model precedent, pricing usually kicks in once the identity is revealed. Second, data handling: third-party compilations indicate that prompts submitted during this preview period are being retained by the provider, with only a promise that they won't be used for training — and who that provider actually is remains unknown. Running open-source projects or small utilities through it is low-risk; company-internal code and configuration files containing keys are better held back until both the identity and the data terms are clear.
Sources: OpenRouter model page and runtime statistics, QbitAI coding tests, CocoLoop, CellCog technical blog, MiniMax developer platform documentation; verified for launch time, context and output length limits, free pricing, throughput and availability methodology, and M3.1-Flash-Preview's published specifications.