MiniMax launches M3.1 Flash at $0.15 per million input tokens

MiniMax launched M3.1 Flash on October 19, making the model immediately available through its API, M Plan and MiniMax Code. The company lists prices of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached-input tokens.

Commercial access does not yet include the model weights. MiniMax said it would release them later but did not provide a date or license. The company also claims M3.1 Flash cuts agent-task usage costs by 70% from its predecessor and makes online inference nearly three times faster. Its launch post does not disclose the workload mix, model settings or matched test conditions needed to independently validate those comparisons.

MiniMax credits more aggressive sparsity for the speed gains. The technique activates only part of a model for a given request. The company says everyday output can exceed 200 tokens per second, with median speed above 170 tokens per second. Those measurements are vendor-reported.

On business-agent tasks, the independent AutomationBench-AA leaderboard places M3.1 Flash second with a 72.5% score, behind Gemini 4 Argon at 77.5%. The benchmark measures completed objectives in simulated software-as-a-service workflows and assigns zero credit to tasks with guardrail violations. MiniMax separately reports a 76.2 score on BankerToolBench, but the opened benchmark materials did not independently reproduce that number.

For autonomous software development, MiniMax reports a 73.7% score on DeepSWE v1.1. The opened DeepSWE leaderboard did not list M3.1 Flash, leaving the figure as a vendor claim.

In computer-use testing, MiniMax reports scores of 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam. The company says both rank first among Chinese models, but neither opened benchmark source contained a matching M3.1 Flash entry.

MiniMax also reports an 82% score on CyberGym and says internal security specialists observed capabilities in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. CyberGym cautions that individual teams submit their own results and that scores can vary because runs are stochastic. The MiniMax result therefore has not been independently reproduced.