Chinese company MiniMax releases fast and cheap AI model for agents

On 19 October 2026, the Chinese AI company MiniMax released a new language model called M3.1 Flash. It is made for AI agents: AI systems that do not just answer a question, but complete a whole job on their own, such as writing and testing code or carrying out a series of actions on a computer.

Such agents have a clear problem: they are expensive and slow. For a single task, the model often has to think, try something and check the result dozens of times. Each step takes computing, and so time and money. A model that answers faster and more cheaply therefore makes agents usable for many more companies and people. That is what MiniMax is aiming for.

Faster and cheaper

According to MiniMax, M3.1 Flash writes almost three times as fast as its predecessor M3. The company mentions a speed of usually more than 170 tokens per second, with peaks above 200. A token is a piece of text of a few letters; 170 tokens per second comes to more than a hundred words per second. MiniMax says it achieves this by letting a smaller part of the model do the calculating at each step.

The price is also going down. Developers pay 0.15 dollars per million tokens they send to the model and 0.60 dollars per million tokens the model writes back. That is half the normal price of M3. According to MiniMax, the total price of an agent task is even 70 per cent lower than with the predecessor.

Tested anonymously first

How the model was tested is striking. Until early October, an early version ran anonymously under the name Space Bunny on OpenRouter, a website where developers can use many different AI models. According to OpenRouter, Space Bunny was the most used model there for ten days and processed 56.7 trillion tokens. Based on user feedback, the model became faster and made fewer mistakes during that period.

Strong at programming and computer work

MiniMax also published a series of test scores. On a test for automating office work, M3.1 Flash scores higher than Claude Opus 5 from Anthropic and DeepSeek V4.1 Flash, according to the company. On a test with banking tasks, it is said to tie with Opus 5. On a programming test in which the model has to solve real software problems, it scores about as high as the strongest competitors. On a tough test in which the model has to type commands into a computer itself, Opus 5 is still clearly ahead. For operating a computer and on a difficult test for agents, MiniMax calls its model the best Chinese model.

All those scores come from MiniMax itself. Independent researchers will now test the model themselves, and those results will show how it really compares with the competition.

Developers can now use the model through MiniMax itself, including in the programming app MiniMax Code, and through OpenRouter, where it is free for the first three days. MiniMax also wants to release the model files soon, so that anyone can run it on their own hardware, just as Meta recently did with its programming model Muse Spark Open. A more powerful Pro version will follow later.

What does this mean for the road to the singularity?

How smart an AI model is counts for only half. For AI agents to really do everyday work, they also have to be fast and affordable. An agent that works on its own for an hour carries out thousands of steps, and each step costs money. Every time the price per step halves, tasks become feasible that were too expensive before.

MiniMax shows that a fast and cheap model can also perform strongly on complex work. What is more, Chinese companies such as MiniMax and DeepSeek are pushing each other on with ever cheaper models, often with public model files. That competition pushes down the price for everyone. If independent tests confirm the figures, the moment comes closer when a team of AI agents can get through a whole day of programming or office work for a few euros.

What is the technological singularity? Read the explanation →