Chinese AI company MiniMax released M3.1 Flash on Oct. 19, a low-cost model it says cuts overall usage costs for agent tasks by 70% compared with its predecessor and runs online inference nearly three times faster.
MiniMax also confirmed that Space Bunny, an anonymous model that it says processed more than 200 trillion tokens while it was being tested under that name, was an early version of M3.1 Flash. Tokens are the small chunks of text that models read and write, and the unit most AI services charge by.
The model is available through MiniMax’s API, its M Plan subscription and its MiniMax Code app. MiniMax says the model weights will be released later, and that a larger Pro model “will also arrive later.”
Price and speed
Pricing is flat: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, according to MiniMax. The model router OpenRouter lists the same prices, and says a free launch period on its service ended Oct. 22.
OpenRouter describes M3.1 Flash as an open-weight mixture-of-experts model with 23 billion active parameters out of 428 billion in total. A mixture-of-experts model only uses part of its network for each piece of text, which keeps it cheaper to run. OpenRouter lists a context window of one million tokens.
MiniMax says it made the model faster through “more aggressive sparsity.” Output speeds can top 200 tokens per second in everyday use, with median speeds above 170, the company says.
The scores
MiniMax reports 73.7% on DeepSWE v1.1, a software engineering benchmark, and 76.2 on BankerToolBench, which tests multi-step financial work. It also reports 61.7 on OSWorld V2, which tests operating a computer across apps, and 29.6 on Agents’ Last Exam, which measures professional work tasks. MiniMax says the model ranks first among Chinese models on OSWorld V2 and Agents’ Last Exam.
For cybersecurity, the company reports 82% on CyberGym, a test of finding and confirming software vulnerabilities. It says its internal security experts found the model strong at code auditing, patching and penetration testing. These are all MiniMax’s own results.
Outside testers paint a mixed picture. VentureBeat reported that the model beat DeepSeek and Anthropic’s Opus 5 on a key business benchmark at lower cost, but that independent testing found a high hallucination rate.
Benchmark results listed on OpenRouter from Artificial Analysis, an independent testing firm, show why.
On Artificial Analysis’s AA-Omniscience knowledge test, M3.1 Flash answered 45.1% of questions correctly but had a non-hallucination rate of 6.3%. That means it usually gave a confident wrong answer instead of saying it didn’t know. The same listing shows 25.8% on Terminal-Bench 4.0, a command-line test.
MiniMax pitches the model at agent work in business software. The company says M3.1 Flash can complete enterprise workflows across sales, marketing, customer service, finance and HR tools from start to finish. It calls Flash part of its “ongoing exploration of the balance between efficiency and capability.”
Source: MiniMax, “MiniMax M3.1 Flash: Redefining Agent Performance and Cost”
Additional sources: OpenRouter, “MiniMax M3.1 Flash – API Pricing & Benchmarks”; VentureBeat, “MiniMax's new M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark at lower cost, but with a high hallucination rate”
