MiniMax Just Launched M3.1 Flash — 3x Faster, 70% Cheaper, and It Was 'Space Bunny' All Along

 Chinese AI company MiniMax officially launched M3.1 Flash on October 19 — and confirmed a juicy secret: the mysterious "Space Bunny" model that secretly topped OpenRouter's leaderboard for 10 straight days was an early version of M3.1 Flash all along.

 What's new

M3.1 Flash is MiniMax's new flagship "Flash" reasoning model, built for agentic work — the kind of AI that chains dozens of model calls together to finish a task. The headline numbers are hard to ignore: inference is nearly 3x faster than its predecessor M3, and the cost of running agentic tasks is down 70%.

Pricing is flat and aggressive: $0.15 per million input tokens and $0.60 per million output tokens. That input price matches DeepSeek's V4.1-Flash-1017 exactly — China's AI price war officially has no floor in sight.

The model ships with a 1-million-token context window, adjustable reasoning across five effort levels, and text, image, and video input. In MiniMax's own testing, it scored 61.7 on OSWorld V2 (a computer-use benchmark, where the company claims no Chinese model has done better), 82% on CyberGym, and 73.7% on DeepSWE v1.

The stealth twist

The "Space Bunny" reveal is the story-within-the-story. Operating anonymously, the model processed over 200 trillion tokens and was OpenRouter's number-one model for 10 consecutive days. It quietly left the platform on October 5, then returned on October 19 as M3.1 Flash — which OpenRouter is offering free for 72 hours. Stealth-testing in public is becoming a genuine launch strategy in this industry. This isn't even MiniMax's first stealth play — back in September, the company's MM3.1-Flash-Preview quietly surfaced on OpenRouter the same way.

Availability

M3.1 Flash is live now via MiniMax's API, its M Plan subscription, and the MiniMax Code app. Open weights land on October 23, and MiniMax says a larger "Pro" model will follow.

Why it matters

AI agents are the year's biggest cost driver — every task can mean dozens of model calls. A model that runs 3x faster while cutting agent costs by 70% doesn't just save money; it makes entirely new kinds of always-on agents economically possible. In a week where speed is the new battleground — from Inception's 1,000-tokens-per-second Mercury 3 to this — MiniMax just fired a serious shot.

One honest caveat: these benchmarks are company-reported and haven't been independently verified yet. And the new model isn't perfect — independent testing noted it tends to answer every question rather than admit uncertainty.

Verdict

Not a frontier breakthrough, but a brutal efficiency play. Faster, dramatically cheaper, and already battle-tested in the open under a fake name. If you're building with AI agents, this is the model to watch this week.