Model Drop

MiniMax M3.1 Flash

October 19, 2026 3 min read

MiniMax M3.1 Flash, which MiniMax released on October 19, 2026, is a cheap, fast model for coding and office agents that spent its preview running anonymously as Space Bunny, priced at $0.15 / $0.60 per million tokens.

Model: MiniMax M3.1 Flash (MiniMax-M3.1-Flash)
Maker: MiniMax
Ship date: October 19, 2026
Pricing: $0.15 / $0.60 per M tokens
Context: 1M tokens
Weights: Open (MiniMax says weights come later)
Where you can use it:
- MiniMax Code, MiniMax's coding app
- MiniMax API and the M Plan
When to reach for it:
- Cheap, fast agents for office and coding work
- Self-hosting once MiniMax posts the weights
How it compares:
- Ahead of DeepSeek V4.1 Flash on office agents
- Trails Claude Opus 5 on coding and computer use
What makes it different:
- Ran anonymously as Space Bunny before launch
- 170+ tokens a second typical output
Early reactions: Early X posts call it a beast on automation for the price; skeptics flag pricier cache hits and possible benchmark tuning.
Read more: MiniMax M3.1 Flash: Redefining Agent Performance and Cost

The benchmarks

MiniMax compared M3.1 Flash with DeepSeek V4.1 Flash and its 1017 update, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. Its table names no effort level, but its cost charts put the 73.7% DeepSWE and 29.6% Agents' Last Exam scores at max effort. Harnesses vary by row (Codex for Terminal-Bench 4.0, mini-SWE-agent for DeepSWE, Claude Code for NL2Repo and CyberGym), Agents' Last Exam is the official full-set leaderboard, and blank cells in MiniMax's table are left out.

MiniMax's numbers

AutomationBench v1.0.6 (business automations): 62.2%. +1.7 over DeepSeek V4.1 Flash 1017 (60.5%), +7.4 over DeepSeek V4.1 Flash (54.8%), +15.5 over Kimi K3 (46.7%), +13.4 over GLM-5.3 (48.8%), +16.4 over GPT-5.6 Sol (45.8%), +11.9 over Claude Opus 5 (50.3%).
BankerToolBench (banking analyst work): 76.2. +2.5 over DeepSeek V4.1 Flash (73.7), +7.1 over Kimi K3 (69.1), +1.7 over GLM-5.3 (74.5), +1.6 over GPT-5.6 Sol (74.6), ties Claude Opus 5 (76.2).
Terminal-Bench 4.0 (multi-hour terminal work): 38.1%. −4.2 under DeepSeek V4.1 Flash 1017 (42.3%), +6.9 over DeepSeek V4.1 Flash (31.2%), +25.5 over Kimi K3 (12.6%), +0.2 over GLM-5.3 (37.9%), −1.8 under GPT-5.6 Sol (39.9%), −13.7 under Claude Opus 5 (51.8%).
OSWorld 2.0 (using a computer): 61.7%. +7.4 over DeepSeek V4.1 Flash (54.3%), +3.4 over Kimi K3 (58.3%), −4.0 under GPT-5.6 Sol (65.7%), −8.9 under Claude Opus 5 (70.6%).
Agents' Last Exam (real professional tasks): 29.6%. +1.3 over Kimi K3 (28.3%), −1.0 under GPT-5.6 Sol (30.6%), −2.6 under Claude Opus 5 (32.2%).
DeepSWE v1.1 (fixing real repositories): 73.7%. −3.3 under DeepSeek V4.1 Flash 1017 (77%), −0.5 under DeepSeek V4.1 Flash (74.2%), +6.2 over Kimi K3 (67.5%), +6.8 over GLM-5.3 (66.9%), +0.7 over GPT-5.6 Sol (73%), −0.3 under Claude Opus 5 (74%).
NL2Repo (building a repo from a spec): 61.2%. −7.0 under DeepSeek V4.1 Flash 1017 (68.2%), −4.2 under DeepSeek V4.1 Flash (65.4%), +3.2 over Kimi K3 (58%), +3.2 over GLM-5.3 (58%), +4.4 over GPT-5.6 Sol (56.8%), −14.1 under Claude Opus 5 (75.3%).
CyberGym (finding security bugs): 82.2%. +2.2 over Kimi K3 (80%), −2.3 under GLM-5.3 (84.5%), −1.4 under GPT-5.6 Sol (83.6%).

Independent numbers

Artificial Analysis hasn't published an Intelligence Index score for M3.1 Flash yet.

First impressions

The model is a few hours old, so these are day-one reactions from X. Hacker News has no launch thread yet.

The positives

OpenRouter, where M3.1 Flash ran anonymously as Space Bunny, says it was the site's most-used model for 10 straight days.

That's real usage before anyone knew whose model it was, and it's free on OpenRouter for the first 72 hours.

Chubby (@kimmonismus), who tracks every lab release, called the benchmarks very impressive given that DeepSeek V4.1 Flash is only days old.

The fair comparison is DeepSeek's cheap Flash line, and MiniMax's own table backs him up on office agents more than on coding.

JUMPERZ (@jumperz), a builder, read MiniMax's cost chart the same way: even the low setting beats GPT-5.6 at max on business automations.

If that holds outside MiniMax's chart, this is the model for repetitive browser and back-office work.

The negatives

stridell (@emilstridell), a developer, did the pricing math against DeepSeek.

MiniMax prices cached input at $0.015 per million, and if he's right about DeepSeek's cache rate, agent loops that reread the same context are where the bill tilts back toward DeepSeek.

StudioYebisu (@studio_yebisu), a Japanese AI watcher, said the demos look strong on vision but the numbers probably don't beat DeepSeek's 1017 update on raw intelligence, and the launch smells a bit like benchmark tuning, as Muse Wonder did.

There's no Artificial Analysis score yet to settle it.