MiniMax Launches M3.1 Flash, a Low-Cost Model With a One-Million-Token Window

MiniMax, the Chinese AI developer, launched its M3.1 Flash model on Monday, October 19, 2026. The model is available through MiniMax’s own API and subscription plans. OpenRouter lists it with a release date of October 19 and a free tier. The Decoder reports that MiniMax describes it as a reasoning model that accepts text and image inputs, with a context window of up to one million tokens.

Pricing is the headline figure. The Decoder lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. OpenRouter also shows a free tier, which is separate from the standard API rates and is the easiest way for developers to test the model without a bill.

The benchmark figures come mostly from Artificial Analysis. The Decoder reports an Intelligence Index score of 42, up from 29 for MiniMax’s earlier M3 model and just below the 45 recorded for DeepSeek’s V4.1 Flash 1017. The article also cites 48 per cent on Humanity’s Last Exam and an estimated cost of about $0.22 per task on the index, compared with $0.26 for the DeepSeek model. MiniMax says the new model runs three times faster than M3.

The same data flags a weakness. The Decoder cites a 93 per cent hallucination rate, which it attributes to the model attempting nearly every question rather than declining when it is unsure. A low price per token does not help a buyer who needs factual reliability, so the accuracy figure should be weighed alongside cost.

The launch also has an unusual backstory. MiniMax says Space Bunny, a model that circulated without a named owner and ranked near the top of OpenCode’s token-usage charts, was an early version of M3.1 Flash, and that it processed more than 200 trillion tokens while anonymous. That claim comes from MiniMax’s own announcement. AI Polix, citing TokenPost, reports that Space Bunny first appeared on OpenRouter in late September, but the token total has not been verified independently.

Open weights are the other open question. The Decoder reports that MiniMax plans to release the weights on October 23. AI Polix, by contrast, reports that the weights have not been published and that MiniMax has not given a date. Until MiniMax publishes the weights or a release note, the model should be treated as hosted-only, and the October 23 date should be considered unconfirmed.

Why it matters: a one-million-token window at these prices puts pressure on rivals. DeepSeek’s V4.1 Flash 1017 has a similar cost per task on the same index, so the price race between Chinese labs is narrowing on cost rather than on capability. The lead on the index is small, and figures from a single benchmark provider will move as more testers run the model.

What is still unknown includes the final weight release date, independent accuracy testing beyond the published index, and how the model performs on long real-world documents at the full context length. Developers who want to estimate how much text they are sending can measure drafts with the free word counter at /en/word-counter before running tests.