MiniMax Releases M3.1 Flash, Cutting Agent Task Costs 70%

The cost and speed gains matter more than the benchmark placement: cheaper, faster inference directly lowers the unit economics of running agent workloads, which is where token consumption compounds. The OSWorld V2 result positions MiniMax competitively among Chinese labs, but first among domestic models is a narrower claim than a global one. The pending open-weight release is the real swing factor — it would let the cost advantage travel beyond MiniMax's own API and M Plan surfaces. Whether that release arrives on the stated terms, and whether pricing at this level holds once rivals respond, is the open question.