'Space Bunny' unmasked as MiniMax M3.1 Flash, a $0.60-output agent model with weights coming 'later'

The developer behind 'Space Bunny Alpha', a model that had been on OpenRouter under a hidden name since late September, has finally been identified. China's MiniMax said in a blog post that the model was its own MiniMax M3.1 Flash (MarkTechPost reported the news on October 19). MiniMax says the model processed more than 20 quadrillion tokens during its anonymous period. M3.1 Flash is built for agents, and MiniMax says its performance is close to the frontier. Even so, output costs only $0.60 per million tokens. But the weights haven't been released yet; MiniMax says they will come 'later'. Nearly all of the performance figures also come from MiniMax's own testing.

Clues left during three anonymous weeks

On October 1, we reported that OpenRouter was hosting a hidden model from an unknown developer. It had a 1 million-token context window and a maximum output of 524,288 tokens, and its free period had been extended to October 5. The model was removed from the listing on October 5, around 16:00–19:00 UTC. No developer claimed it.

At the time, the closest match on specs was MiniMax's M3.1-Flash-Preview. On September 27, Beijing time, MiniMax had released that model only inside its own coding tool, MiniMax Code. It also had a 1 million-token context window. We called MiniMax "the most likely candidate, with moderate confidence." Now that MiniMax has confirmed it, the guess turns out to be correct.

On OpenRouter, it's now common for a developer to release a model without its name, test it under real-world load, and reveal itself afterward. Developers use the model without knowing who made it, so how much it gets used becomes the verdict. MiniMax went through this stage before its formal announcement.

Where $0.60 output sits

Here is how $0.60 per million output tokens compares with other recently released models.

Top-tier models cost much more. Claude Opus 5.5 charges $20 for output, GPT-6 Sol $10, and StepFun's Step 5 Preview $2.70. That puts M3.1 Flash at somewhere between a few percent and about a fifth of their price. At the cheap end, GPT-6 Luna and Claude Haiku 5.5 (for prompts of 100,000 tokens or fewer) both charge $0.50 for output. Xiaomi's MiMo-V2.6-Flash costs even less at $0.28, and NaiveAI's Naive-N0.5-Flash costs $0.40. M3.1 Flash is not the cheapest model available.

The price still matters because MiniMax is pitching M3.1 Flash as a near-frontier agent model, not as a 'cheap small model'. In setups that give each subagent its own model, a high-performance model does the planning and many cheaper models run in parallel to carry out the work. In those setups, the per-token price of the models doing the work largely sets the total cost. A $0.60 model with a 1 million-token context and a high output limit is aimed squarely at that role.

Still, per-token prices alone don't tell you what a model will cost. In October, Claude Haiku 5.5 had a low per-token price but used about 162,000 tokens per task, about three times as many as GPT-6 Luna. How many tokens M3.1 Flash uses per task isn't known yet. That number will decide how cheap $0.60 really is.

Weights 'later', and the numbers are its own

M3.1 Flash comes with two important caveats.

The first is the weights. MiniMax says it will release them later, but they aren't out yet. In the same week, DeepSeek released the weights for V4.1-Flash-1017 under the MIT license, and StepFun released the weights for Step 5 Preview on October 15, as it had promised. While its Chinese rivals deliver weights on schedule, M3.1 Flash is for now available only through the API. When the weights will come out, and under what license, is the next thing to watch.

The second is the performance numbers. Most of the comparison figures cited in reporting are self-reported by MiniMax. Artificial Analysis has created a listing for the model, but no third party has yet done a broad check of the agent abilities MiniMax is promoting. There are recent cases of vendor numbers falling short in independent testing. Xiaomi said MiMo-V2.6-Pro was 'on par with Opus 5', but it scored 5 points lower on the Artificial Analysis index. On Terminal-Bench 4.0, Anthropic reported 39.2% for Haiku 5.5, while Artificial Analysis measured 33%.

The figure of 'more than 20 quadrillion tokens' during the anonymous period also comes from MiniMax, and it isn't clear what was counted (input plus output, or over what period). For comparison, OpenRouter's figures from mid-September put weekly volume across all Chinese-made models at about 67 trillion tokens. 20 quadrillion is far larger than that. Until the counting method is known, the number is best read as MiniMax's claim.

What to check next

Three things will decide how M3.1 Flash is judged. First, when the weights are released and under what license. Second, whether third parties such as Artificial Analysis and Vals AI test it on agent tasks and publish the cost and token count per task. Third, how much developers kept using it after the anonymous period, which will show up in OpenRouter's rankings in the coming weeks.

On October 8, MiniMax shares fell 13.74% in Hong Kong to HK$207.2. While China's leading open-weight developers compete on releasing weights and on price, MiniMax first tested its model's performance anonymously and then started competing on price. For now, the weights aren't out, and the performance numbers are MiniMax's own. Whether $0.60 is really cheap remains to be seen.