On Artificial Analysis’s AA-Omniscience test, the hallucination rate is the share of questions a model did not get right where it gave a wrong answer instead of declining. It measures whether a model knows when to say it does not know, which is separate from how often it is right.
What happened
MiniMax released M3.1 Flash, an agentic model with a one-million-token context window at $0.15 per million input tokens and $0.60 per million output. VentureBeat's report puts it at 42 on the Artificial Analysis Intelligence Index, up from 29 for M3, with the average cost per index task falling from about $0.51 to $0.22.
The same report has two numbers from AA-Omniscience, Artificial Analysis's knowledge test, that look like they disagree. M3.1 Flash answered 56% of the questions correctly, ahead of DeepSeek V4.1 Flash at 43% and GLM-5.3 Flash at 28%. Its hallucination rate was 93%. Both are true, and the gap between them is the useful part.
What 93% does not mean
It does not mean 93% of its answers were wrong. Artificial Analysis's benchmarking methodology scores each response as correct, wrong, or an abstention, and the index "subtracts points for hallucinated responses" while it "keeps abstentions neutral". The hallucination rate looks only at the questions a model did not get right, and asks how many of those it answered wrongly rather than declining.
Put the two numbers together. M3.1 Flash got 56 of every 100 questions right, which leaves 44. At a 93% hallucination rate, about 41 of those 44 were confident wrong answers, and about 3 were "I don't know". In other words, it almost never declines. When it does not know, it guesses, and the guess reads exactly like a right answer.
Accuracy and hallucination rate are different questions
A model can raise its accuracy just by refusing less, and its hallucination rate rises at the same time. That is why a single "accuracy" figure in a launch post tells you less than it seems to.
When a model that guesses is fine
A high hallucination rate matters in proportion to how much of the answer comes from the model's memory. It matters least when something else supplies or checks the facts:
- The source is in the prompt. Summarising a document, extracting fields from an invoice, answering from retrieved passages. The model is reading, not recalling. See RAG done properly for how to keep it reading.
- The output is checked by machine. Code that has to pass tests, JSON that has to match a schema, a tool call that fails loudly when it is wrong. A guess that fails a check costs you a retry, not a mistake.
- A person reviews every output anyway. First drafts, brainstorming, rewriting your own text.
It matters most when the model answers from memory and nobody checks: facts about people and companies, legal or medical questions, citations, numbers, and any agent that acts on what it believes without looking it up. Those are the cases where "I don't know" was the right answer, and a guessing model will almost never give it.
Before you switch, test it on your own work
Run the same set against the model you use now. The comparison you care about is not a leaderboard position. It is how many wrong answers each model would have handed you on your own questions.
What is still open
Several of the strongest claims in the launch, including the business-workflow results against Claude Opus 5 and GPT-5.6 Sol, are MiniMax's own comparisons, and VentureBeat notes the evaluation settings differ. The weights are proprietary at launch, with a release planned for October 23, and the licence terms have not been published. If you plan to self-host, wait for the licence before you plan around it.
The story is in the October 19 briefing. For how independent scores feed our picks, see the leaderboard. To compare per-token prices, read LLM pricing compared, and for a wider checklist, how to evaluate an AI tool's reliability.
Explore curated tools related to this guide: