MiniMax M3.1 Flash vs. MiniMax M3: One is cheaper, but not faster or better

MiniMax launched its new M3.1 Flash model on October 19, 2026, claiming a 70% reduction in overall costs for agentic tasks and up to three times faster online inference compared to its predecessor, M3. The company lists Flash's pricing at $0.15 per million input tokens and $0.60 per million output tokens, which is half the discounted rates of M3 at $0.30 and $1.20 for requests with up to 512K input tokens.

Independent testing by Artificial Analysis scored M3.1 Flash at 72.5% on AutomationBench-AA, measuring the percent of task objectives completed without guardrail violations. This score differs from the percentage of entire tasks completed on MiniMax's own AutomationBench leaderboard.

To test the models, I ran three smaller agent tasks through OpenRouter, a service that lets users test multiple LLM models. The tests were run with default settings and capped turns to measure performance. Flash completed all three tasks with a score of 5 out of 6, while M3 achieved the same score but missed one case on the second bug.

Flash's approach involved reading and editing more files and consuming more tokens, resulting in a 23% lower estimated cost. In contrast, M3 was faster overall, finishing the tests in 24 turns compared to Flash's 17 turns.

For the incident investigation test, both models scored 10 out of 10, correctly identifying the cause of an outage and listing all affected services. However, Flash used fewer input tokens and had an estimated cost about 68% lower than M3. The long context test also yielded no results for either model, as neither produced a complete answer within the given token limits.