MiniMax has released M3.1 Flash, a model aimed squarely at enterprise workflows, software development and computer use, and positioned on price. The company claims roughly 3x the speed of its predecessor M3 and a 70% cut in overall agent usage cost. Independent testing from Artificial Analysis largely confirms the efficiency story and adds a warning that matters for anyone deploying it unsupervised.
Cost and Capability
At maximum reasoning effort, M3.1 Flash scores 42 on the Intelligence Index, up from 29 for M3, with average cost per task falling from roughly $0.51 to $0.22. That puts it level with Z.ai’s GLM-5.3 Flash ($0.28 per task) and just ahead of Gemini 3.8 Flash ($1.24), within reach of DeepSeek V4.1 Flash-1017, which scores 45 at $0.26.
The reasoning gains are pronounced: 48% on Humanity’s Last Exam, up from 39% and ahead of DeepSeek’s 43%, and 22% on CritPt, up from just 4%. The model also generated roughly 189 million output tokens during the index run, against 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash — a saving that has nothing to do with list price.
MiniMax’s own figures claim 62.2 on Automation-Bench v1.0.6, ahead of DeepSeek V4.1 Flash-1017 at 60.5, 76.2 on BankerToolBench (tying Claude Opus 5), 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, and 82.2 on CyberGym. API pricing starts at $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015. Open weights are due on October 23.
The Hallucination Problem
On Artificial Analysis’s AA-Omniscience benchmark, M3.1 Flash answered 56% of factual questions correctly — ahead of DeepSeek V4.1 Flash-1017 at 43% and GLM-5.3 Flash at 28%. Yet it attempted nearly every question and produced a 93% hallucination rate, yielding an overall score of just 16.
The metric needs care: it measures the share of non-correct responses that were outright wrong, not the share of all answers. The real finding is that when M3.1 Flash does not know, it almost never declines or hedges — it invents. For agents that update records, reconcile financial documents, or research customers unsupervised, confident fabrication is a materially worse failure mode than a refusal, and it argues for retrieval, guardrails, or human review rather than raw benchmark scores.
M3.1 Flash remains behind Anthropic’s and OpenAI’s frontier systems on overall index performance. But with frontier-adjacent task economics and open weights arriving, it is another data point in the same direction: capable, inexpensive models are spreading into workloads that previously assumed an expensive model was the only option.