MiniMax M3.1 Flash: MiniMax’s Flash model, 3x faster and 70% cheaper for agents

MiniMax reports 61.7 on OSWorld V2 and 76.2 on BankerToolBench. Reasoning stays on for every request, with developers choosing one of five effort levels from Low to Max. List pricing is $0.150/M input, $0.015/M cached input, $0.600/M output, and the free launch period ended Oct 22, 2026. The model circulated anonymously before launch. An anonymous model called Space Bunny turned up on OpenRouter's model routing service in late September. Independent tests have pushed back on the accuracy side: what both reports agree on is directional: M3.1 Flash is faster and cheaper, and it appears to hallucinate more often than the higher-priced models it is being compared against. Open weights are also unsettled, since one listing calls it an open-weight model released October 20, 2026, while another lists download availability as not available. Agent workloads are where this matters. The release takes aim at the cost of running AI agents, which often need dozens of model calls to finish a single job. Founders should judge a model by cost per completed task, not list price per token. The 70% figure compares against MiniMax's own predecessor, so teams should rerun it on their own workloads, with error rates included, before moving production agents.