MiniMax unmasks Space Bunny as its new M3.1 Flash model

For ten straight days in late September, the most-used model on OpenRouter had no company name attached. On October 19, MiniMax claimed it: "Space Bunny," the anonymous model that OpenRouter says processed 56.7 trillion tokens while sitting at No. 1, was an early build of MiniMax M3.1 Flash, which the Chinese company launched the same day through its API, its M Plan subscription and MiniMax Code.

The price is the point. M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens, half of what MiniMax charges for M3 and exactly what DeepSeek asks for V4.1 Flash during off-peak hours. In China's race to the bottom on fast, cheap models, MiniMax has chosen to match its strongest rival rather than undercut it.

What $0.60 per million output tokens buys

MiniMax's launch post promises "a generational leap in agentic tasks, coding, Computer Use and cybersecurity, while making online inference nearly three times faster and reducing usage costs by 70%." The model takes text and image input, has a 1-million-token context window, and the company has not disclosed its parameter count.

Independent numbers are more modest. Artificial Analysis scored M3.1 Flash at 42 on its Intelligence Index at launch, level with GLM-5.3-Flash, one point above Gemini 3.8 Flash and three behind DeepSeek V4.1 Flash-1017. Its cost per task came to about $0.22, against $0.51 for M3: a cut of roughly 57%, not 70%. MiniMax's own benchmark table concedes ground too, showing DeepSeek ahead on DeepSWE (77.0 versus 73.7) and Terminal Bench 4.0 (42.3 versus 38.1).

Open weights are promised but not yet out. MiniMax's post on X said they "will arrive on Oct 23," while its blog says only "later." Nothing is on Hugging Face yet, and no license has been published. MiniMax also says Space Bunny handled more than 200 trillion tokens across platforms, a figure it has not broken down; OpenRouter's 56.7 trillion is the only third-party count, and that traffic was served free.

Fast on paper, slow on Plotly's exam

Speed claims run into a familiar catch. Plotly's data-analysis benchmark scored M3.1 Flash 41 out of 43, tying Anthropic's Opus 5.5, but at a median of three minutes per question, against 36 seconds for the Space Bunny preview. "It seems to use A LOT of tokens and it's the slowest model that we've tested for the exam; it's slow because of the number of tokens it uses, not the inference speed," Plotly co-founder Chris Parmer wrote on Hacker News.

Reliability is the bigger flag. On Artificial Analysis's knowledge test, the model "attempts almost every question, resulting in a hallucination rate of 93%." VentureBeat's Carl Franzen drew the lesson: "High benchmark intelligence alone does not guarantee that a model will reliably identify its own limitations."

Why MiniMax matched DeepSeek's price

MiniMax has shipped at a fast clip. MiniMax-M1, released in June 2025, offered open weights and a 1-million-token context; M2 and M2.5 followed, and M3 arrived in June under a community license. The company listed in Hong Kong in January, raising about $619 million.

Growth has not closed the gap to profit. First-half 2026 revenue rose 283% to $116.6 million, but the adjusted net loss was $293.0 million, with $1.32 billion in cash. MiniMax shares fell after DeepSeek launched V4.1-Flash in September, SiliconANGLE reported, citing Bloomberg. Matching that price is defensive as much as ambitious, and The Decoder concluded that "neither model wins across the board."

CEO Yan Junjie framed the constraint in August: "Intelligence can scale almost without limit; energy and compute cannot." Friday's promised weight release, and the license attached to it, will show how open MiniMax intends to stay. A larger Pro model has been announced without a date.