MiniMax M3.1 Flash (Space Bunny) Released
Read the Portuguese translation →
Author: Data Science in your pocket · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: MiniMax M3.1 Flash (Space Bunny) Released · Channel: Data Science in your pocket · 02:06 · Watch on YouTube
Topics: AI agents, AI evaluation and benchmarks, Cybersecurity and software security, Model inference and optimization, AI models
Summary
Section titled “Summary”MiniMax confirmed that Space Bunny, an anonymous model that processed over 200 trillion tokens, was an early version of MiniMax M3.1 Flash. The release targets agentic tasks, coding, computer use, and cybersecurity, with nearly three times faster inference and a 70% reduction in usage cost. It scores highly on several agentic and cybersecurity benchmarks, and the video highlights its low cost per task. It does not lead in every area: Claude Opus 5 remains ahead on terminal work, long refactoring sessions, and pure reasoning depth.
Key points
Section titled “Key points”- MiniMax confirmed Space Bunny was an early version of MiniMax M3.1 Flash.
- Space Bunny reportedly processed over 200 trillion tokens while its identity was unknown.
- M3.1 Flash is aimed at agentic tasks, coding, computer use, and cybersecurity.
- Inference is nearly three times faster, peaking above 200 tokens per second.
- Usage cost is reported to be 70% lower.
- The model scores 62.8 on Automation Bench, 76.2 on Banker Tool Bench, and 73.7 on Deep SWE.
- It scores 82.2 on Cyber Gym, within about two points of the leader.
- The video says M3.1 can reach the top of Automation Bench at roughly 4¢ per task.
- Pricing is 15¢ per million input tokens and 60¢ per million output tokens, with cached input costing a fraction of a cent.
- Claude Opus 5 still leads on terminal work, long refactoring sessions, and pure reasoning depth.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Space Bunny revealed | MiniMax confirms the anonymous Space Bunny model was an early version of M3.1 Flash. |
| 00:32 | Speed, cost, and benchmarks | The video covers M3.1 Flash’s speed, reduced usage cost, and results on agentic benchmarks. |
| 01:04 | Cybersecurity and cost per task | M3.1 Flash’s Cyber Gym score and low Automation Bench cost are highlighted alongside its pricing. |
| 01:35 | Trade-offs and conclusion | The video notes where Claude Opus 5 still leads and frames M3.1 Flash as a cost-effective model for agentic work. |
Quotes
Section titled “Quotes”The wait is over. — 00:00
It is nearly three times faster at inference — 00:32
But the real story is cost per task. — 01:04
It is trying to be the smartest model per dollar on agentic work. — 01:35
Concepts and entities
Section titled “Concepts and entities”MiniMax M3.1 Flash, Space Bunny, MiniMax, Claude Opus 5, AutomationBench, Banker Tool Bench, Deep SWE, Cyber Gym
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Remember Space Bunny? An anonymous model that quietly processed over 200 trillion tokens while the whole community guessed who was behind it. The wait is over. Minimax has confirmed that Space Bunny was an early version of its new release, Minimax M3.1 Flash. And this is not a minor point update. M3.1 Flash targets the exact workloads that matter in 2026, agentic tasks,
00:32 coding, computer use, and cybersecurity. It is nearly three times faster at inference, peaking above 200 tokens per second, while usage cost drops by 70%. On Minimax’s own agentic benchmarks, it sits at or near the top, 62.8 on automation bench, 76.2 on banker tool bench, tied with Claude Opus 5, and 73.7
01:04 on Deep SWE. Cybersecurity is the sleeper strength. 82.2 on Cyber Gym, within about two points of the leader. But the real story is cost per task. M3.1 maxes out automation bench at roughly 4¢ per task, where frontier models charge close to a dollar for the same score. Pricing is flat and simple. 15¢ per
01:35 million input tokens, 60¢ per million output, and a fraction of a cent for cached input. To be fair, it does not win everything. Claude Opus 5 still leads on terminal work, long refactoring sessions, and pure reasoning depth. M3.1 Flash is not trying to be the smartest model on Earth. It is trying to be the smartest model per dollar on agentic work. Space Bunny hop through
02:06 200 trillion tokens in stealth. Now we get to use it.