By Hamza Faisal
The most talked-about AI model of October wasn't announced by anyone. For weeks, developers watched an anonymous "Space Bunny" rocket past established models on public leaderboards, processing trillions of tokens without a company name attached. On Monday, Chinese startup MiniMax solved the mystery: Space Bunny was an early version of its new flagship, M3.1 Flash — and it was never really hiding at all.
The mystery model had a name all along
MiniMax confirmed the connection in a launch post featuring a carrot-shaped rocket and the message "Welcome back, Space Bunny." According to the company and financial newswire Cailian Press, the stealth model processed more than 200 trillion tokens while operating anonymously — the largest stealth run ever recorded. OpenRouter, the routing platform that hosted it, said Space Bunny held its number-one spot for ten consecutive days, processing 56.7 trillion tokens on that platform alone.
OpenRouter added that feedback gathered during the anonymous run directly improved the final model: time to first token dropped 38 percent, throughput rose 45 percent, and the error rate was cut in half. As a launch gesture, M3.1 Flash is free on OpenRouter for 72 hours.
Three times the speed, a fraction of the cost
MiniMax describes M3.1 Flash as delivering "generational" improvements in agentic tasks, programming, and cybersecurity. The headline numbers: online inference speed nearly tripled, and usage costs fell 70 percent compared with its predecessor.
The model is priced aggressively — $0.15 per million input tokens and $0.60 per million output tokens, with cached input cheaper still. Under the hood, it packs 428 billion total parameters with just 23 billion active at a time, a million-token context window, multimodal input covering text, image, and video, and adjustable "thinking" depth across five effort levels.
On the benchmarks that matter for agent workloads, the company claims M3.1 Flash leads globally on at least three widely used tests — Automation Bench, Banker ToolBench, and LOCA-Bench-256K — making a strong case for the best performance-per-dollar in agent workflows. Independent analyses note that DeepSeek's rival V4.1 Flash model still edges it on some coding cost curves, but concede MiniMax sits firmly in second place.
China's efficiency playbook keeps working
The launch extends a remarkable run for Chinese AI labs. Despite operating with what analysts estimate is fewer than 100,000 H100-equivalent GPUs — a fraction of the compute available to American labs — MiniMax and DeepSeek keep pushing the frontier curve, a trend that has made distillation-based explanations of their progress increasingly hard to sustain.
Investors noticed: MiniMax shares, listed on the Hong Kong Stock Exchange after a blockbuster January debut, rose 1.45 percent to close at HK$182.1 on the news.
M3.1 Flash lands in a crowded month for model releases — DeepSeek, doubleAI, and Microsoft's own 2.8-trillion-parameter effort are all jostling for attention. But the Space Bunny episode may be the more lasting story: a reminder that in 2026, the most effective model launch isn't a press conference. It's letting the product win anonymously first.
By Hamza Faisal
