MiniMax launches M3.1 Flash AI model, claims 70% lower cost

Chinese artificial intelligence company MiniMax released a new AI model, M3.1 Flash, on Monday, October 19, 2026. The company says the model is better than its previous generation at agentic tasks, coding, operating a computer and cybersecurity work, while answering nearly three times faster and costing about 70% less to use. It is priced at $0.15 per million input tokens and $0.60 per million output tokens, and is available through MiniMax's API and its subscription products.

A token is the small piece of text, roughly a part of a word, that AI models read and write; companies charge for AI by the million tokens. At $0.15 per million input tokens, M3.1 Flash is in the cheapest tier of capable models, at about half the per-token price of MiniMax's own M3 and in line with the off-peak rates DeepSeek charges for its V4.1 Flash model. The race among Chinese labs to cut prices is pushing down the cost of building AI products everywhere.

The model is aimed at agentic work. An AI agent is a system that does not just answer a question but carries out a series of steps, such as reading files, writing and testing code, or clicking through a computer screen to finish a task. Agents use far more tokens than a chat, so a lower price per token matters more for them. That is why MiniMax has stressed both speed and cost alongside its benchmark scores.

The claims need care. The benchmark figures come from the company, and benchmarks measure narrow tasks under test conditions, not reliability in daily use. The independent results from Artificial Analysis are a reminder of this: a model can score well on agent tests and still give confident wrong answers, a problem known as hallucination. For anyone building on the model, those two sets of numbers point to the same advice the testers give: check its output before relying on it.

For Indian developers and start-ups, many of which build products on top of models they rent through APIs, falling prices change the arithmetic of AI products, especially for Indian-language and high-volume services where cost per user must stay low. Choosing a model from any foreign provider also raises questions about where data is processed and how it is protected, which Indian firms must weigh under the country's data protection rules.

Short-term: Developers can try M3.1 Flash now through MiniMax's API and partner platforms. Competing providers face more pressure to cut prices on their own fast, low-cost models.

Long-term: If open weights are released as promised, companies will be able to run the model on their own hardware, which can lower costs further and keep data in-house. Continued price cuts could make AI agents affordable for smaller firms and for services in markets such as India.

Who is affected: Software developers and start-ups building AI agents, companies comparing AI model costs, rival AI labs in China and the United States, and users of the products built on these models.

Questions and Answers