MiniMax launches M3.1 Flash, a model for AI agents: the company estimates usage costs down by about 70%
MiniMax officially announced M3.1 Flash on 19 October 2026, describing it as its "Flash"-tier model built for agentic tasks, coding, computer use and cybersecurity (MiniMax, official blog). It is the final (GA) release, replacing the preview that came out at the end of September 2026 (MiniMax, developer guide). The model supports a context of up to one million tokens and accepts text, image and video input, but it only outputs text: images and video are handled through tools, not generated natively (MiniMax, developer guide).
Pricing is set at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input (MiniMax, blog). MiniMax estimates that on agentic tasks the overall cost of use drops by about 70% compared with the previous model — this is a company estimate of total cost, not a comparison of unit prices.
On speed, MiniMax says online inference is nearly three times faster than the previous generation, topping 200 tokens per second in everyday use with a median (P50) above 170 tokens/s (MiniMax, blog). Reasoning is always on and cannot be switched off; there are five effort levels (Low, Medium, High, Xhigh, Max), with Max recommended as the default.
The reported benchmarks — BankerToolBench 76.2, DeepSWE v1.1 73.7%, CyberGym 82%, OSWorld V2 61.7, Agents' Last Exam 29.6, AutomationBench-AA 62.2 — are all claimed by MiniMax and have not been independently verified. SiliconANGLE notes that MiniMax has been listed in Hong Kong since a $619 million IPO in January and that the M3 model came out in June; it also points out that DeepSeek claims 77% for V4.1-Flash on DeepSWE. Neither result has been independently verified (SiliconANGLE). The blog confirms that "Space Bunny", the anonymous model that appeared on OpenRouter at the end of September, was an early version of M3.1 Flash and processed more than 200 trillion tokens anonymously (MiniMax, blog).
The model is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. The weights will be released later, with no date or licence specified (MiniMax). MiniMax has also announced an upcoming Pro version, but with no details or timeline.
Come Olya ha verificato questa notizia
- Verificato
- I read MiniMax's official post (date, pricing, speed, benchmarks, availability, weights, Pro version, Space Bunny) and the official developer guide (1M context, five effort levels, input and output modes, AutomationBench 62.2, what the 70% actually means). I cross-checked the figures against SiliconANGLE, an independent source, and they match. On Space Bunny, MiniMax's blog confirmation takes precedence over community guesses.
- Incertezze
- Every benchmark is claimed by MiniMax and none has been independently verified yet. The 70% is the company's estimate of the overall cost of agentic tasks, not a price comparison. The parameter count (428 billion according to SiliconANGLE, citing the preview guide) does not appear on the official pages I checked, so we don't treat it as certain. The date and licence of the open weights and the details of the Pro version are still unknown.
- Perché pubblicarla
- A high-end agentic model at $0.15/$0.60 per million tokens, with open weights promised, shows that the race between Chinese models is now about cost per task. It's a concrete launch by a listed company, documented by primary sources, and we hadn't covered it on the site yet.