MiniMax M3.1 Flash lowers AI costs without dispelling doubts

MiniMax has launched M3.1 Flash, a reasoning model rated 42 points on the Artificial Analysis Intelligence Index. Its price per task drops, but its tendency to hallucinate remains high.

MiniMax has launched M3.1 Flash, a multimodal reasoning model that accepts text and images, with a context window of up to one million tokens. According to measurements cited by Artificial Analysis, it scores 42 points on the Artificial Analysis Intelligence Index, while bringing the average cost per task down to about 0.22 dollars. The price drop is clear, but it does not solve everything: the model retains a high hallucination rate, which immediately matters for professional uses.

Lower prices, but a bill that depends on actual work

MiniMax displays an API rate of 0.15 dollars per million input tokens, 0.60 dollars per million output tokens, and 0.015 dollars for cached inputs. The company also says that M3.1 Flash is nearly three times faster than M3. On paper, the promise is clear; in an AI agent, the final cost mainly depends on the number of loops needed to complete a task, not just the unit rate.

Artificial Analysis provides a more precise framing: the average cost per task falls to 0.22 dollars, compared to 0.51 dollars for M3. The gain thus exists, but it is rather around 57% in this benchmark, not the 70% claimed by MiniMax. In other words, the model cuts the bill, but not necessarily as much as the commercial communication suggests.

Solid scores, especially on complex reasoning

On the tests mentioned in the sources, M3.1 Flash progresses significantly compared to M3. It rises from 29 to 42 points on the Artificial Analysis Intelligence Index, scores 48% on Humanity’s Last Exam, 22% on CritPt, and 56% on AA-Omniscience. In this last test, it surpasses DeepSeek V4.1 Flash 1017, rated at 43% correct answers.

The picture remains nuanced, however. DeepSeek keeps the advantage on the Artificial Analysis Intelligence Index with 45 points, and also leads on CritPt with 24%, compared to 22% for MiniMax. On coding uses, M3.1 Flash shows 73.7% on DeepSWE v1.1, a level very close to its direct rivals. The Chinese race for cheap models increasingly looks like a war of pennies with millimeter rankings. Excel tables, however, do not suffer from hallucinations.

The real sensitive point: trust in the answer

The most troubling figure comes from the AA-Omniscience test. The model answers correctly 56% of factual questions, but 93% of its wrong answers are assertions rather than abstentions. In other words, it tries almost everything, even when it should remain silent. For a general public chatbot, this may be acceptable; for a research, analysis, or business assistance tool, it is more delicate.

MiniMax also presents results on OSWorld V2, with 61.7 points, and on CyberGym, with 82%. These figures suggest interesting capabilities in software interaction and security, but the sources specify that exact test conditions are not detailed. They should therefore be read as signals, not as definitive proof.

Wide availability, and a very monitored openness strategy

The model is already available via MiniMax's API, and the company plans to release the weights as open weights on October 23, according to sources. A desktop application, MiniMax Code, also accompanies the launch. With a context of one million tokens, the model can absorb large codebases or long documents without immediate splitting, making it relevant for technical assistance or synthesis uses.

One important unknown remains: the license was not yet published at the time of the sources, and MiniMax also mentions a Pro version to come with no firm date. In this context, teams considering a real deployment would do better to test their own use cases, with verification, safeguards, and human review. The price drops. Vigilance remains at the same level.

By the numbers

  • 42 points — Artificial Analysis Intelligence Index for M3.1 Flash.
  • 0.22 dollars — average cost per task according to Artificial Analysis.
  • 1 million tokens — context window announced by MiniMax.
  • 56% — correct answers on AA-Omniscience.
  • October 23 — planned date for open weights availability.

To read