make it so cheap that it is “invisible” to the user

DeepSeek is the tip of the spear when we talk about artificial intelligence from China. Not only does it have great performance, but Microsoft itself has raised the alarm by pointing out that its policy is allowing it to gain users in markets where others like OpenAI have a more difficult time. Other companies like Tencent or Alibaba are taking giant steps in the fight for AI, and a few days ago ByteDance -TikTok- presented a Seedance 2.0 that is impressive… and is already giving you headaches.

But the big ones are not the only ones, and with China focused on the development of robotics and AI, we must talk about other smaller ‘players’. Zhipu AI and MiniMax are two of the “tigers” that, in just a few years, have raised hundreds of millions of dollars and that have models that have a radically different philosophy from that of OpenAI and other Western giants.

Their models are sold as life companions, tools that people can use on a daily basis without worrying about the Price. And, within that discourse, MiniMax has just launched the M2.5, a model that wants to become a “digital employee” and that its managers have classified as its first “frontier model” so cheap that it is not worth measuring the price.

AI too cheap to worry about price

M2.5 is now official and, as reported by the South China Morning Post, MiniMax did not want to waste the opportunity to launch it in a hectic week for the AI ​​industry in China. Technically, M2.5 is an LLM – large language model – that can handle about 230 billion total parameters, but only uses 10 billion per token. Being a Mixture of Experts system, each call only involves the experts directly necessary to resolve the request.

Bringing the figure down to earth, that means that it is a capable model, but by user request does not use its full potentialwhich implies low inference costs and very low prices for users. Those responsible for it claim that the price is just one dollar per hour of continuous operation, spending 100 tokens per second. This means that you can have an “agent” working continuously for that entire time at a price between 10 and 20 times lower than other models such as Opus, Gemini 3 Pro or GPT-5.

Such an aggressive policy makes M2.5 a model “too cheap to quantify,” according to those responsible, facilitating mass adoption because the user can stop optimizing each order he gives to the AI. That phrase “too cheap to meter” is a nod to the historic comment in which it was pointed out that electricity from nuclear energy would be too cheap to measure.

And something important is that M2.5 is not a simple chatbot. It is available on platforms such as Ollama, HuggingFace, ModelScope in China or GitHub, and MiniMax itself points out that 30% of the company’s internal tasks are already carried out by M2.5 itself. Furthermore, 80% of new code is generated by the model. That is, it is more optimized for working alone than for chatting. This code created by code thing is not exclusive to M2.5, and Codex and Opus is also in this boat.

The model has already been tested and, although in some tasks it achieves notable results, especially compared to other models open-weightits score is far from that of the closed models. In the company’s own internal results, it managed to double the score of the previous model, the M2.1, but as SCMP points out, these internal benchmark scores are difficult to verify independently.

Internal benchmark in coding | Image: MiniMax

But, in the end, whether it is more or less capable compared to other models, MiniMax M2.5 is one more example of the strategy that China is pushing with artificial intelligence. While the United States is striving to demonstrate that it has increasingly powerful and capable proprietary models, AI is in a narrative in which it aims to promote cheaper and more useful models for the user.

This implies not only that they have a good performance/price ratio, but also that they can run on everyday devices without enormous computing power. And now that certain Chinese companies will supposedly be able to get their hands on some of NVIDIA’s best GPUs to train AI, the boost to that strategy could be notable.

Images | MiniMax (edited)

In WorldOfSoftware | There is another race equally important as the one for chips to win AI and in that China takes the lead