MiniMax has released the M3.1 Flash reasoning model, achieving a score of 42 points on the Artificial Analysis Intelligence Index at less than half the cost of its predecessor. The new model scores 48 percent on Humanity's Last Exam and 22 percent on the CritPt physics reasoning benchmark, while recording a factual accuracy of 56 percent on the AA-Omniscience test.
The model handles up to one million context tokens and accepts text and image inputs. API pricing is set at $0.15 per million input tokens and $0.60 per million output tokens, with cached inputs priced at $0.015, bringing the average cost per task down to $0.22. MiniMax runs three times faster than the M3 model, though Artificial Analysis notes a 93 percent hallucination rate because it attempts to answer all test questions instead of declining when uncertain.
MiniMax offers the API on its platform and provides a desktop application named MiniMax Code, with open weights scheduled for release on October 23. In benchmark comparisons, the model trades blows with Deepseek V4.1 Flash 1017, leading in factual accuracy and complex reasoning exams while remaining slightly cheaper per task.
Why it matters: This release highlights how aggressive pricing and improved reasoning capabilities are driving intense competition among Chinese artificial intelligence developers.