The 3x and 70% are company claims

The IT Home page is stamped 2026/10/19 20:20:45. It reports MiniMax’s announcement the same day: M3.1 Flash is live. The page says Space Bunny, which had anonymously processed more than 200T tokens, was an early version of this model. The page calls it a new Flash model. The speed and the cost are the company’s own figures: online inference is nearly 3 times faster, and the cost of use is 70% lower. There is no independent retest in this piece.

The scores on the same page are also numbers from the company’s description. BankerToolBench is printed as 76.2. DeepSWE v1.1 is printed as 73.7%. OSWorld V2 is printed as 61.7, and Agents' Last Exam (ALE) as 29.6. The page says those last two rank first among domestic models. AutomationBench-AA is only called a front-runner, with no score. The page also says that, on real workflows across SaaS systems, it can fully handle problems in sales, marketing, operations, support, finance, and HR. The page does not name the comparison models, does not say how many samples each test used, and does not say who measured the scores. “First” and “front-runner” are therefore not places on an outside leaderboard.

The unit prices are renminbi printed on the page, not a conversion. Input is 1 yuan per million tokens, output is 4 yuan per million tokens, and cache reads are 0.1 yuan per million tokens. The 70% cut is written for agentic tasks and is compared with the previous model. The previous unit prices are not printed. The three rates and the 70% cut are not one price list that can be subtracted.

The page says the model is already live on the MiniMax API, MPlan, and MiniMax Code, that the weights will open later, and that a Pro version aimed at a higher capability ceiling will also arrive later. Later has no date. The page has a further security evaluation. This edition does not use it.