MiniMax M3.1 Flash triples speed, cuts agent cost 70%

MiniMax frames MiniMax M3.1 Flash as a direct answer to the cost problem in agentic AI. Agent tasks chain many tool calls together, so per-token prices compound quickly; MiniMax says it cut the cost of running a whole agent task by 70% versus its prior model, and set API input at $0.15 per 1M tokens. It also claims median output speed above 170 tokens per second, about triple the prior model's online inference speed.

The company links the model to a real-world test. It says 'Space Bunny,' an anonymously operated service that processed more than 200 trillion tokens, was an early version of MiniMax M3.1 Flash — an unusually concrete claim about pre-release deployment. MiniMax also asserts first place among Chinese models on OSWorld V2 and Agents' Last Exam, and cites scores of 76.2 on BankerToolBench and 73.7% on DeepSWE v1.1.

Independent evaluator Artificial Analysis scored it 42 on its Artificial Analysis Intelligence Index, against a median of 13 for models in the same price band. It also flagged heavy token consumption — about 180 million output tokens in the overall evaluation versus a median of about 99 million — so low unit prices do not automatically mean low bills. MiniMax says it will publish the model weights later and add a 'Pro' model afterward.