The most-used coding model has no name

OpenCode publishes its own usage data, and for the past week the top row has belonged to a model with no company attached. Space Bunny ran 58 trillion tokens through the coding tool over seven days, ahead of DeepSeek V4.1 Flash at 34T and Meta's Muse Spark at 31T. Nobody has claimed it.

The model showed up on OpenRouter and OpenCode on September 23 as Space Bunny Alpha: one million tokens of context, image and video input, tool calling, five reasoning effort levels, and a price of zero during the preview. OpenRouter's listing describes an anonymous model run by a third-party provider that chose to stay unnamed. The developer has not disclosed the model family, the parameter count, or the knowledge cutoff. "Alpha" is doing a lot of work in that name.

The adoption curve is the part worth staring at. Space Bunny's first day in OpenCode's daily table was September 23, at 1.5T tokens. The next day it did 6.8T. The day after that, 9.0T, and it has held the top daily spot since. A guide written after the model's first week put it at the top of OpenRouter's daily ranking too, three days after launch, ahead of DeepSeek V4.1 Flash and GLM 5.3 Flash. Whatever this thing is, people did not need convincing.

OpenCode also breaks usage down by model author, and that view holds the strangest number. On September 22, the "Other" bucket where unattributed models land carried 1.4 percent of tokens. By October 3 it carried 40.9 percent. Meta sits at 23.4 percent, DeepSeek at 23.2 percent, OpenAI at 0.3 percent. The largest supplier of coding-agent tokens in this dataset is a residue category for models with no name on them.

The user counts tell a second story. On October 3, Space Bunny ran on about 41,000 unique OpenCode users, second to DeepSeek's 46,000. Fewer people, far more tokens each. That is the signature of long, heavy sessions rather than a wide base of casual tries, which fits a free 1M-context model exactly the way you would expect.

Should you trust it? There are no official benchmarks. The numbers circulating are unofficial slices, roughly 82 percent on a 60-question GPQA cut and 75 percent on MMLU-Pro, promising and unfalsifiable in equal measure. Independent tokenizer studies point at MiniMax, which shipped an M3.1 Flash preview on September 27 and has never linked itself to the bunny. Treat that as a hypothesis until someone claims the model.

There is a precedent hiding in plain sight. The previous stealth listing on OpenRouter, Ox Alpha, turned out to be GLM-5.3-Flash, and the free part of that arrangement ended. A codename is a temporary thing. The model ID you wire in this week is not the one you will be billed for in December.

None of this makes Space Bunny a scam. Free tokens from an unknown seller are still free tokens, and 58 trillion of them a week means the model does real work. It does mean leaderboard position has quietly changed meaning. A ranking that used to measure whose model was best now also measures who can fund the most free capacity, and the data cannot tell you which of those you are looking at. Keep the model ID in configuration with a named fallback beside it, and let the preview earn its place. The bill arrives when the mask comes off.