MiniMax M3.1 Flash Revealed as Space Bunny: 3x Faster, 70% Cheaper, Open Weights October 23
MiniMax officially launched M3.1 Flash on October 19 and confirmed what had been one of the more curious open secrets on the router circuit: Space Bunny 🐰, the anonymous model that had quietly processed over 200 trillion tokens on OpenRouter without a lab attribution, was an earlier version of it.
The model ships at $0.15 per million input tokens and $0.60 per million output tokens, with cached reads at $0.015. Open weights are promised for October 23. On OpenRouter, it ran free for the first 72 hours after launch.
Capabilities and Claims
M3.1 Flash carries a 1M token context window and accepts text, image, and video input. Thinking effort is adjustable — not fixed at a preset level. MiniMax positions the model for agentic tasks, software development, computer use, and cybersecurity.
The headline performance claims are vendor-reported:
- 3x faster online inference vs the M3 generation
- 70% lower cost on agent-type workflows
- Enterprise workflow tasks priced at 0.18–0.30 yuan per completion (~$0.03–0.04)
On MiniMax’s internal benchmarks, it ranks among the top Chinese models on Automation-Bench v1.0.6 and OSWorld 2.0, without specifying a numeric placement.
Benchmark Numbers
Third-party benchmark results, sourced from the vLLM team’s day-zero evaluation coverage:
| Benchmark | Score |
|---|---|
| HLE | 48% |
| BabyVision (with tools) | 91.0 |
| CharXiv RQ | 82.5 |
| AA-Omniscience | 56% |
| BankerToolBench | 76.15 |
| CritPt (scientific reasoning) | 22.3% |
These figures are from MiniMax’s official benchmark data compiled by third parties; independent harness results are not yet available.
The Space Bunny Story
Space Bunny Alpha appeared on OpenRouter with no lab attribution in an arrangement MiniMax used as a production-scale stealth beta. OpenRouter confirmed it was MiniMax’s model and noted that MiniMax shipped speed and error-rate fixes during its anonymous run. The model processed over 200 trillion tokens in that period.
This approach — ship to production at scale, fix in place, name it later — differs from standard model launch conventions, where betas are labelled and staged.
Pricing Context
At $0.60 per million output tokens, M3.1 Flash sits below the flash-tier price band of most major Western labs. GPT-6 Luna prices at $0.50 per million output tokens, the closest peer by tier. DeepSeek V4.1 Flash is significantly cheaper at roughly $0.029 per task on LiveBench’s costing data, though it operates in a different regime as a fully open-weight model.
No SWE-bench Verified or tau-bench results have been published for M3.1 Flash. Standard agentic harness evaluations will determine where it actually lands against the frontier.