MiniMax M3.1 Flash Tops Agent Benchmark, Offers Flat All-Day Pricing

MiniMax also confirmed that Space Bunny, an anonymous model that had drawn attention in developer communities, was M3.1 Flash. OpenRouter disclosed that the anonymous model had been the platform's most-used model for 10 consecutive days and processed 56.7 trillion tokens; MiniMax said cumulative processing during testing and use exceeded 200 trillion tokens.

According to benchmark data released by MiniMax, M3.1 Flash ranked first in Automation-Bench v1.0.6, a set built by SaaS automation platform Zapier with 600 real enterprise business-process tasks across sales, marketing, operations, customer service, finance and HR. It tied Opus 5 for first on BankerToolBench at 76.2% and scored 82.2% on CyberGym. In coding, DeepSeek remained ahead: on DeepSWE v1.1, DeepSeek scored 77.0% and M3.1 Flash 73.7%, close to Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%. MiniMax said M3.1 Flash was second only to DeepSeek on coding cost-effectiveness.

As of Oct. 20, M3.1 Flash's intelligence index on the Artificial Analysis leaderboard reached 43, up from 29 for the previous M3, while its task-weighted average model cost fell to $0.22 from $0.51. MiniMax priced the model at 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan for cache reads. That is roughly half the input and output prices of the previous M3 and about 75% lower for cache reads. Compared with DeepSeek V4.1 Flash-1017, MiniMax said its input and output prices are half of DeepSeek's peak rates and equal to its idle rates, with no peak or off-peak distinction all day.

Developer Hassan said on X that when he asked M3.1 Flash and Claude Opus 5.5 to build a launch landing page for a hardware product, MiniMax's design satisfied him, the bills were $0.014 and $1.22, and MiniMax was about four times faster, according to Leiphone.

Leiphone ran its own test: given a hand-drawn campus map and a building dimension table, the models had to build from scratch an interactive 3D campus in a browser, including layout and conflict checks, Blender modeling scripts, editable source files, GLB export and a Three.js page supporting building selection, detail views, camera positioning and simulated time. MiniMax Code with M3.1 Flash took 55.4 minutes; OpenCode with DeepSeek V4.1 Flash took about 121 minutes. Both completed the main deliverables. At official API prices checked on Oct. 20, MiniMax's model cost was about 6.50 yuan and DeepSeek's about 6.52 yuan at peak rates. MiniMax's cache hit rate reached 98.1%.

DeepSeek's page leaned toward building-data display with a left-hand list, a right-hand detail panel and a bottom time slider. MiniMax put building numbers, site information and view switching more prominently, and clicking a building produced an orange highlight outline and floating labels that made the link between the scene and building data clearer. Both kept editable source files, and the exported GLB models passed checks for six building numbers.

MiniMax also cited ALE-v1 Overall results: across 152 tasks, M3.1 Flash had a 29.6% pass rate at an estimated model cost of about 739 yuan, while GPT-5.6 Sol had about 30.6% at about 5,000 yuan, a gap of about 1 percentage point in pass rate and nearly seven times in cost. The 29.6% pass rate also indicates that M3.1 Flash is far from replacing flagship models outright, and failed retries and human checks still count in real use. MiniMax said it will later launch a Pro version aimed at a higher intelligence ceiling.

Editor's Summary

MiniMax released M3.1 Flash two days after DeepSeek's V4.1 Flash-1017, claiming first place on Automation-Bench's complex agent tasks and competitive all-day pricing. In Leiphone's 3D campus test, the MiniMax setup finished in 55.4 minutes versus about 121 minutes for DeepSeek at nearly identical model cost. The launch gives developers a lower-cost option for agent workloads, though M3.1 Flash's 29.6% ALE pass rate shows it is not a full replacement for flagship models.