MiniMax released M3.1 Flash on October 19, pitching it as a faster, cheaper model for coding and AI agents. Independent testing found a higher score on a broad intelligence benchmark, but also raised a practical concern: when the model got a factual question wrong, it often gave an incorrect answer rather than admitting uncertainty.

The model is available through MiniMax’s API, M Plan subscription and MiniMax Code. MiniMax’s announcement lists these prices:

  • $0.15 per million input tokens
  • $0.60 per million output tokens

What the price does, and does not, cover

MiniMax says M3.1 Flash runs nearly three times faster than its predecessor and cuts usage costs for agent tasks by 70%. Those are company claims, not independent guarantees for every workload.

At the listed rates, one million uncached input tokens and one million output tokens would cost $0.75 in model usage. That gives developers a useful starting point for budgeting. It does not include other costs, such as infrastructure or human review.

The low token price is clear. The harder question for businesses is how to catch a confidently wrong answer before it affects a workflow.

What independent testing found

VentureBeat’s report on Artificial Analysis testing says M3.1 Flash scored 42 on the Intelligence Index, up from 29 for MiniMax’s previous M3 model. The index combines results from multiple evaluations. It does not guarantee performance in a particular company’s workflow.

On Artificial Analysis’s AA-Omniscience factual-knowledge test, M3.1 Flash answered 56% of questions correctly. Its reported hallucination rate was 93%. That does not mean 93% of all its answers were false. The figure describes the share of non-correct responses that were outright wrong, rather than partial answers or refusals.

Cheap automation still needs checking

Lower token prices may make frequent model use more affordable for routine coding and multi-step tasks. But if a system supplies a wrong answer and updates records or informs a business decision without checks, the result could be extra work or a costly mistake.

For US businesses considering an AI agent, the practical test is to compare cost and performance on the tasks it will actually handle. The benchmark score is one data point. Teams still need to decide where a person or another verification step must remain in the process.