MiniMax M3.1 Flash cuts task costs as benchmark scores rise
Illustration, not documentary evidence of the event.
MiniMax has released M3.1 Flash, a reasoning model that improves on its predecessor’s benchmark performance at less than half the reported cost per task, according to the-decoder.com. It scores 42 on the Artificial Analysis Intelligence Index, up from M3’s 29, with a reported benchmark cost of about $0.22 per task.
The API is live, with pricing of $0.15 per million input tokens and $0.60 per million output tokens. The model accepts text and images and supports up to one million tokens of context. MiniMax plans to release open weights on October 23. Its claim that the model runs three times faster than M3 comes from the company.
For teams choosing a model, the results suggest a candidate for evaluation rather than an automatic replacement. Against Deepseek V4.1 Flash 1017, MiniMax scores lower on the overall Intelligence Index, 42 versus 45, but higher on Humanity’s Last Exam, 48 percent versus 43 percent. Those mixed results make the relevant decision narrower: which model handles the work your application actually needs?
Uncertainty handling deserves a separate check. The report gives M3.1 Flash a 56 percent correct-answer score on AA-Omniscience, while also citing Artificial Analysis’s 93 percent hallucination rate and its explanation that the model attempts every test question instead of declining when uncertain. These benchmark figures do not establish an error rate for your application. Test representative questions with known answers alongside questions the model should decline, then assess correct answers and unsupported responses separately.
The cost comparison also needs a workload-specific check. MiniMax’s reported $0.22 per task versus Deepseek’s $0.26 reflects lower output-token use across the benchmark, not a fixed charge for every request. Run the same representative tasks through both models, record billed token use and review answer quality. For an application that needs reliable answers, compare cost per acceptable result, including any review or retries your workflow requires, before treating the lower benchmark bill as a saving.