MiniMax ships M3.1 Flash, a faster and cheaper agentic model
The Chinese lab says the new model cuts agentic task costs by 70% against its predecessor, with output speeds above 200 tokens a second.
MiniMax released M3.1 Flash on 19 October, pitched as a faster, cheaper replacement for its agentic model. It’s live now through the MiniMax API, the M Plan subscription and MiniMax Code.
The headline numbers are speed and cost. MiniMax says output runs at over 200 tokens a second, with median throughput above 170 TPS, and that online inference is “nearly three times faster” than its predecessor. Pricing is $0.15 per million input tokens and $0.60 per million output, which the company says cuts the cost of agentic tasks by 70%. On benchmarks MiniMax published itself, the model tops its own AutomationBench-AA ranking and scores 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2. Open weights aren’t out: MiniMax says only that they’ll follow “later,” with no date given.
Why it matters: if you’re running coding or tool-use agents through an API rather than self-hosting, a 70% price cut with a speed increase on the same provider is the kind of change that reshuffles which model is cheapest for a given job.
The caveat: every number here comes from MiniMax’s own post, there’s no independent benchmark yet, and the weights needed to verify or self-host it aren’t available.