MiniMax M3.1 Flash: $0.15 per million input tokens, and it guesses

MiniMax just made a frontier-adjacent agent model very cheap. Before you route work to it, read the hallucination number.

MiniMax released M3.1 Flash on Monday, October 19, priced at $0.15 per million input tokens and $0.60 per million output tokens. It is live in the MiniMax API, its M Plan subscription and MiniMax Code.

What MiniMax shipped, and what it claims

The price list, from MiniMax:

$0.15 in, $0.60 out, $0.015 for cached input, per million tokens.

MiniMax says agentic tasks cost 70% less to run than on its previous model, and that output reaches more than 200 tokens per second in everyday use, with a median above 170.

Its own benchmark table claims 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. MiniMax says the model ranks first among Chinese models on OSWorld V2 and Agents' Last Exam. VentureBeat reports that the BankerToolBench score ties Claude Opus 5 in MiniMax's results.

The weights are not out yet. MiniMax says they will follow later, and VentureBeat reports a planned release on October 23, with no license published so far.

The independent numbers are more mixed

Artificial Analysis lists M3.1 Flash at 43 on its Intelligence Index, fourth of the 173 models on the page, with a 1 million token context window. VentureBeat reported a score of 42 at the model's maximum reasoning setting, up from 29 for MiniMax's M3.

That is a real jump in capability per dollar.

Two other numbers cut the other way. Artificial Analysis measures output on MiniMax's API at about 116 tokens per second, well under the speeds MiniMax quotes. And VentureBeat reports a 93% hallucination rate on Artificial Analysis's AA-Omniscience test.

That 93% needs reading carefully. VentureBeat says it is the share of the model's non-correct answers that were wrong guesses rather than an admission it did not know. It is not 93% of all answers. VentureBeat reports that Artificial Analysis found an "exceptionally high tendency to generate incorrect answers rather than acknowledge uncertainty."

It still means this: when M3.1 Flash does not know, it rarely says so.

Use it where a wrong answer gets caught

For agent loops that check their own work (code that has to compile, tests that have to pass, forms that either submit or fail), a confident wrong guess gets caught. The low price wins there.

For anything that ends in a fact a person relies on, a hallucination that sounds sure of itself costs more than the tokens you saved. Research summaries, support answers and anything legal or financial belong in that bucket.

So split the work. Put M3.1 Flash on the high-volume, verifiable steps. Keep a model that will say "I don't know" on the steps that end in a claim.

Run your own evals before you switch. Most of the benchmark table is MiniMax's own.

What to watch

October 23, and the license that comes with the weights. A restrictive license changes who can self-host it.

Independent scores beyond Artificial Analysis, especially on agent tasks.

Price is the headline. Accuracy is the bill.