MiniMax M3.1 Flash costs $0.15 in and $0.60 out per million tokens, but its hallucination rate is 93%. Give it a way to say unknown

A cheap model that never says "I don't know" is fine, as long as your workflow asks the question differently.

What happened

MiniMax released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index, 13 points above its predecessor M3. The Decoder reports it ties GLM-5.3-Flash and trails DeepSeek V4.1 Flash 1017 at 45. It answers 56% of questions correctly on the AA-Omniscience factual test, ahead of DeepSeek's 43%.

API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015. The model handles up to one million tokens of context and accepts text and images. Because it uses fewer output tokens than rivals, its cost per task on the index is about $0.22. MiniMax says it scores 73.7% on the DeepSWE v1.1 coding benchmark, runs nearly three times faster than M3 and costs 70% less for agentic tasks.

The weak spot: a 93% hallucination rate, which The Decoder attributes to Artificial Analysis, because the model tries to answer every test question instead of declining when it is uncertain. MiniMax plans to release open weights on October 23.

My take

A high hallucination rate does not mean you avoid the model. It means you design around how it fails. This one does not say "I don't know" on its own, so your workflow has to make that easy.

How I would use a model like this in a client automation:

• Give it the source text and ask it to extract, not recall. Do not ask it facts from memory
• Make "unknown" or null a valid answer in your output schema, and say so in the prompt
• Validate outputs against the record: does that email, ID or amount actually appear in the input?
• Route anything it marks uncertain, or anything that fails validation, to a human or a stronger model

At $0.60 per million output tokens, it is a candidate for high volume steps like tagging, summarizing call notes or classifying tickets. Run it beside your current model on a sample of real data first, and measure how often it invents something before it touches a live CRM.