MiniMax M3.1 Flash: Space Bunny Unmasked at $0.15 per 1M
MiniMax M3.1 Flash is the stealth "Space Bunny" model that topped OpenRouter. Price: $0.15 input and $0.60 output per 1M tokens, free for a few days, weights promised later.
- Official pricing per 1M tokens: $0.15 input, $0.60 output, $0.015 cached input; MiniMax claims about 3x faster speed and 70% lower cost than its previous model on agent workloads.
- OpenRouter said Space Bunny sat at #1 for 10 days and served about 56.7T tokens; MiniMax says the early build handled more than 200T tokens across platforms.
- Polymarket (~06:11 UTC Oct 20): MiniMax at 2.4% to have the best Chinese model on the Arena text leaderboard at month end, far behind Xiaomi at 59.5%.
in this block
MiniMax M3.1 Flash is out, and the mystery model that topped OpenRouter for days finally has a name. On October 19, MiniMax said the anonymous "Space Bunny" was an early build of MiniMax M3.1 Flash, priced it at $0.15 per million input tokens and $0.60 per million output tokens, and promised open weights later. Coding agents switched it on for free within hours.
What actually happened
MiniMax published a launch post for MiniMax M3.1 Flash on Monday, and Reuters picked up the release the same day. The key reveal was the Space Bunny connection. For more than a week, developers had been using a stealth model on OpenRouter and in the OpenCode coding agent without knowing who built it. OpenRouter said it held the top spot for 10 straight days, and OpenCode said it ran #1 there for 12 days with about 106T tokens processed.
The model is now back under its real name, and both platforms are offering it free for a limited window: OpenRouter said 72 hours, OpenCode said three days. Cline, Ollama Cloud and several other tools also flipped it on. After that, the regular price kicks in, so early users should note when their free window closes, roughly around October 22.
The specs, and who says so
MiniMax says the model streams at more than 200 output tokens per second, with a median above 170. MarkTechPost reported it as a mixture-of-experts model with about 428B total parameters and roughly 23B active per token, a 1M-token context window, and open weights scheduled for October 23. The license for those weights was not disclosed, so do not assume "open" means a permissive license until it ships.
On benchmarks, all numbers are MiniMax's own: 62.2 on Automation-Bench, 73.7 on DeepSWE, 91.0 on BabyVision and 38.1 on Terminal Bench 4.0. That last one is notably below the 51.8 MiniMax lists for Claude Opus 5, which is a nice bit of honesty in a launch chart. MiniMax also said a bigger Pro version is coming later, with no date.
Cheap is the whole pitch
The fun numbers are on cost. One widely shared developer test claimed a single coding task cost about $0.014 with the new Flash model versus $1.22 with Claude Opus 5.5. That is one person's run, not a benchmark, but it shows why agent builders care. When a coding agent loops through hundreds of tool calls, price per token decides whether you can afford to let it think.
Not everyone is sold. Some developers on X said the Flash model felt weaker than rival open models for hard tasks, and the Terminal Bench score backs up the idea that it is a speed-and-price play, not a frontier reasoning model. We saw the same tradeoff with DeepSeek V4.1 Flash 1017 and with Moonshot's Kimi K2.6: cheap Chinese models win usage first, then argue about quality.
What Polymarket thinks
Polymarket's "Best Chinese AI Company end of October?" market resolves on the arena.ai Text Arena overall leaderboard, style control off, checked October 31 at noon ET. At about 06:11 UTC on October 20, MiniMax traded at 0.024, with a bid of 0.022 and an ask of 0.026 on roughly $15K of volume. Xiaomi led at 0.595, Moonshot was at 0.155, Alibaba at 0.10 and DeepSeek at 0.027.
That pricing makes sense. A Flash model built for throughput rarely tops a human-preference leaderboard against flagship releases, and the market is about the single highest-ranked Chinese model, not the most used one. Traders are betting MiniMax wins on usage charts and loses on the Arena, which are two very different scoreboards.
What to do as a reader
If you build with coding agents, the free window is a cheap way to test MiniMax M3.1 Flash on your own tasks. Run the same prompts against your current model, compare cost and failure rate, and keep a log. Do not move production traffic based on a launch chart or one viral cost screenshot.
If you care about open weights, wait for the October 23 drop and read the license before building on it. And if you are trading the leaderboard market, remember it resolves on Arena rank alone. Nothing here is investment advice.
Sources: MiniMax launch post and Polymarket market page.
Not financial advice. DYOR, ser.
