New to this? Read it in simple words
- MiniMax, a Chinese AI company, released a new model called M3.1 Flash.
- It costs half as much per token as MiniMax-M3, the model before it.
- An outside tester ranks it fourth of 173 models.
- But the tester found it often gives wrong answers instead of saying it does not know.
- Token
- A small piece of text, often part of a word. AI companies charge by the number of tokens.
- Hallucination
- When an AI gives a confident answer that is wrong or made up.
- Weights
- The numbers inside a model. When a company releases them, anyone can run the model.
MiniMax is a Chinese AI company. Its models are used by developers who want strong results at a low price. In June it released the weights of MiniMax-M3, the model before this one.
What MiniMax released
M3.1 Flash is built for agents, programs that carry out tasks step by step. It is available through MiniMax’s API, its subscription plan and its coding tool. It reads up to 1 million tokens of text and images at once.
The price is $0.15 per million input tokens and $0.60 per million output tokens. MiniMax-M3 costs $0.30 and $1.20, so per token the new model is half the price, according to MiniMax’s own price list.
MiniMax says the model makes online use “nearly three times faster” and cuts usage costs by 70%. The 70% compares the whole cost of agent tasks, not the price per token.
MiniMax’s claims beside Artificial Analysis’s test of October 19.
What the outside test found
Artificial Analysis, which tests models on the same set of tasks, gives M3.1 Flash 42 points on its Intelligence Index. That puts it fourth of 173 models. MiniMax-M3 scored 29, The Decoder reports.
The tester measured 120.1 tokens per second, close to the average. MiniMax claims more than 200 tokens per second. Running the whole index cost $0.22 per task.
On a test of facts, the model gave a wrong answer instead of declining in 93% of the cases where it did not know, The Decoder reports from Artificial Analysis. “The model tries to answer every test question,” The Decoder wrote.
Claims and open questions
MiniMax reports its own scores on agent tests, for example 61.7 on OSWorld V2, a test of using a computer. No outside group has repeated those tests yet.
Reuters reports that MiniMax claims “generational” improvements in agent tasks, programming and cybersecurity.
“Model weights will be released later,” MiniMax wrote, without a date. The Decoder and MarkTechPost report that the weights are due on October 23. MiniMax has not published the licence.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3Research · October 19, 2026MiniMax-M3.1-Flash: Intelligence, Performance & Price Analysis Artificial Analysis
- 4Research · October 19, 2026MiniMax M3.1 Flash shows China’s AI price race has no floor in sight The Decoder
- 5Research · October 19, 2026China’s MiniMax launches updated AI model M3.1 Flash, claims major improvements Reuters (via AOL)
- 6Research · October 19, 2026MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input MarkTechPost
The speed, the 70% saving and the agent test scores are MiniMax’s claims. The index score, the measured speed and the hallucination rate come from Artificial Analysis.
MiniMax’s new model costs half as much and ranks fourth of 173, but an outside test found a 93% hallucination rate. How we rate the mood