News update: this post was researched and drafted automatically from the sources listed at the end, and published without a human edit. If you spot a mistake, write to me and I will fix it.
MiniMax has released M3.1 Flash, a new reasoning and agentic model. It is a 428 billion parameter mixture of experts model, according to MarkTechPost. The model handles up to one million tokens of context and accepts text and images. MiniMax plans to release the weights as open weights on October 23. The API is live now at platform.minimax.io, priced at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens.
MiniMax says M3.1 Flash is nearly three times faster than its predecessor and cuts overall usage cost for agentic tasks by 70 percent. The Decoder reports that Artificial Analysis puts the model at 42 points on the Intelligence Index, 13 points higher than the previous M3 model, and the cost per task is about $0.22. DeepSeek V4.1 Flash 1017 scores 45 points and costs about $0.26 per task on the same evaluation. So M3.1 Flash does not beat DeepSeek everywhere, but it is cheaper per task because it uses fewer output tokens.
In one line: MiniMax released M3.1 Flash, a 428B parameter mixture of experts model with a 1M token context, open weights due October 23, agentic and coding scores near DeepSeek V4.1 Flash, and a lower cost per task on Artificial Analysis.
Where the model improved
MiniMax reports a set of benchmark numbers. On DeepSWE v1.1, a benchmark for software tasks that require understanding code, developing features, fixing issues, and validating through testing, M3.1 Flash scores 73.7 percent. MiniMax says this places it alongside top models. On AutomationBench-AA it ranks among top performers for real enterprise workflows across SaaS systems. On BankerToolBench it scores 76.2, doing well on multi-step planning, document consistency, financial compliance, and chained tool calls.
The Decoder adds more numbers. On Humanity's Last Exam, M3.1 Flash hits 48 percent, beating both DeepSeek and GLM. DeepSeek manages 43 percent. On CritPt, a research-level physics reasoning benchmark with 71 problems, it scores 22 percent, ahead of GLM-5.3-Flash at 15 percent and Gemini 3.8 Flash at 18 percent, but behind DeepSeek at 24 percent. On AA-Omniscience for factual accuracy, it answers 56 percent correctly, against GLM-5.3-Flash at 28 percent and DeepSeek at 43 percent.
For computer use and cybersecurity, MiniMax reports 61.7 points on OSWorld V2, 29.6 points on Agents' Last Exam, and 82 percent on CyberGym. MiniMax says it ranks first among Chinese models on OSWorld V2 and ALE. The company also says internal security experts found the model strong in code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering.
Price and token efficiency
The flat API price is low. Input costs $0.15 per million tokens, output $0.60, cached input $0.015. The Decoder notes that M3.1 Flash uses about 189 million output tokens across the Artificial Analysis benchmark, compared with 224 million for DeepSeek V4.1 Flash. That brings the cost per task to about $0.22, against $0.26 for DeepSeek. MiniMax claims output speeds of 200 plus tokens per second in everyday use, with median P50 speeds of 170 plus tokens per second.
This matters for teams like mine who run many agentic tasks in a day. A lower token count per task reduces bill shock, not just the unit price. If the model can complete a job in fewer output tokens, the effective cost per task drops even when the headline price is similar.
The catch
Artificial Analysis reports a 93 percent hallucination rate for M3.1 Flash. The reason, according to The Decoder, is that the model tries to answer every test question instead of declining when it is uncertain. That is a serious flag for client delivery work. In a RAG chatbot for a regulated client, a model that rarely says it does not know can produce confident wrong answers. I would not put this in front of a client without a strong evaluation harness that checks refusal behaviour and groundedness.
Open weights are not available yet. MiniMax says the weights will be released on October 23. Until then, we can only evaluate through the API or the MiniMax Code desktop app. The absence of local weights matters for data sensitive clients, because self-hosting is often a requirement.
What I would watch
For my work, agentic coding and RAG chatbots are the two places I would test this model. The DeepSWE number is encouraging for software development workflows, and the lower cost per task is useful for long running agent loops. But I would not swap production models on day one. I would run it through our standard client evaluation set, including cases where the model should refuse, and compare output token counts and latency against our current default. The open weights release on October 23 is the next checkpoint. If the weights drop as promised and the hallucination pattern improves or can be controlled with prompting, this becomes a serious option for internal rollouts and maybe for client pilots.
Nothing to do yet beyond watching and maybe a small API trial. The price is real, the open weights date is close, and the benchmark mix is relevant. The main risk is trust.