MiniMax has officially launched **M3.1 Flash**, a fast and low-cost model aimed at agentic AI. The company says it reaches frontier-level results on key tasks at **one third the cost of M3 and three times the speed**. The announcement on X also confirms that the **model weights will arrive on 23 October**. Here are the numbers and what they mean. More on this topic: TechnoSports MiniMax M3.1 Flash coverage. Key Details at a Glance DetailMiniMax M3.1 FlashPositioningFast, low-cost agentic modelVersus M31/3 the cost, 3x the speedAPI price (flat rate)$0.15 per 1M input tokens, $0.60 per 1M output tokensCache reads$0.015 per 1M tokensWeightsArriving 23 OctoberEarlier versionSpace Bunny, which topped leaderboards with 200T+ tokensHeadline benchmarksAutomation-Bench 62%, ALE 30%, DeepSWE 74%, BabyVision 91% https://twitter.com/MiniMax_AI/status/2112111925975032241 What MiniMax Announced MiniMax says M3.1 Flash achieves frontier-level intelligence across key tasks in three areas: Cowork (Automation-Bench 62%, ALE 30%), coding (DeepSWE 74%) and visual reasoning (BabyVision 91%). It adds that Space Bunny, a model that topped global leaderboards with more than 200 trillion tokens served, was an earlier version of M3.1 Flash. Pricing is a flat rate of $0.15 per million input tokens and $0.60 per million output tokens, with cache reads at $0.015 per million tokens. Access is through the MiniMax platform API, the M Plan and MiniMax Code for desktop. Benchmark Chart MiniMax shared a full benchmark table with the launch post. Here are a few rows, as read from the chart, comparing M3.1 Flash with Opus 5 and GPT-5.6 Sol.

BenchmarkM3.1 FlashOpus 5GPT-5.6 SolAutomation-Bench v1.0.662.250.345.8BankerToolBench76.276.274.6Terminal Bench 4.038.151.839.9DeepSWE v1.173.774.073.0NL2Repo-Bench61.275.356.8CyberGym82.2-83.6BabyVision (w/tools)91.094.188.9HLE w/o tools43.455.649.5 The full chart is shown in the image below, and you can also view the original post on X: MiniMax announcement. The chart shows M3.1 Flash ahead on Automation-Bench and level with rivals on DeepSWE and BankerToolBench, while Opus 5 leads on Terminal Bench, NL2Repo-Bench and Humanity's Last Exam without tools. Treat these as vendor-reported numbers. MiniMax M3.1 Flash benchmark chart, as shared by MiniMax on X. Why the Price Matters At $0.15 and $0.60 per million tokens, M3.1 Flash is priced for high-volume agent workloads, where tokens add up fast. A model that is three times faster and a third of the cost of M3 could let teams run more agent steps for the same budget. Open weights on 23 October will also let developers host the model themselves. Read more about the company on MiniMax and see our AI model coverage. How to Read the Claims Benchmarks from the model maker are a starting point. Run M3.1 Flash on your own tasks, check latency under load and compare it with other low-cost models before moving production work. Earlier, a preview called M3.1-Flash-Preview went live on 27 September inside MiniMax Code, according to DataNorth, so this launch is the wider public release.

https://twitter.com/MiniMax_AI/status/2112150867868541221 Quick Answers How much does MiniMax M3.1 Flash cost? $0.15 per 1M input tokens and $0.60 per 1M output tokens, with cache reads at $0.015 per 1M tokens. When do the weights arrive? On 23 October, per MiniMax. How fast is it? MiniMax says 3x the speed of M3 at one third the cost. Source: MiniMax on X, MiniMax platform, MiniMax on Wikipedia. Who Should Try It Teams building coding agents, workflow automation or document-heavy assistants are the natural users. The flat pricing and cache read discount reward long, repeated prompts, which is how most agents work in practice. Start with a small pilot on your own tasks, compare cost per finished job and not just cost per token, and wait for the open weights on 23 October if you want to self-host.