MiniMax has launched M3.1 Flash, a new reasoning model that achieves 42 points on the Artificial Analysis Intelligence Index. This result represents a 13-point increase over the previous M3 version while reducing the cost per task by more than half. The model supports up to one million tokens of context and processes both text and image inputs.
In performance testing, M3.1 Flash reached 48 percent on the Humanity's Last Exam benchmark and 22 percent on the CritPt physics reasoning test. It also demonstrated 56 percent accuracy on the AA-Omniscience test. Artificial Analysis noted a 93 percent hallucination rate, as the model attempts to answer all questions rather than declining when uncertain.
The model operates three times faster than its predecessor. API pricing is set at $0.15 per million input tokens and $0.60 per million output tokens, with cached inputs costing $0.015 per million. The total cost per task is approximately $0.22, which is lower than the $0.26 cost associated with the Deepseek V4.1 Flash 1017 model.
MiniMax plans to release the model with open weights on October 23. The API is currently available via the company's platform, and the developer also provides a desktop application named MiniMax Code. The model competes directly with Deepseek V4.1 Flash 1017, with both companies pursuing open-weight strategies for their latest releases.
