Chinese AI Firms Unveil Faster, Cheaper Models for Agent Market

Chinese artificial intelligence (AI) companies are releasing new models that work faster and cost less than earlier versions. Chinese AI was already known for being affordable, but these new models aim to cut costs even more and complete tasks faster. They are designed for the AI agent market, where computer systems carry out tasks on their own based on users' goals. Meanwhile, U.S. companies such as OpenAI and Anthropic are continuing to develop new models, making the global AI competition even stronger. This means the "AI development speed control theory," suggested by technology leaders last month, now seems largely pointless.

Chinese AI startup MiniMax released its new model, "M3.1 Flash," on the 19th (local time). It is the successor to M3, which came out in June. The new model cuts the cost of AI agent tasks by 70% and makes the process of generating answers three times faster. It uses a design called "Mixture of Experts (MoE)." This design chooses 23 billion parameters, or about 5% of its total 428 billion parameters, based on each request. The model can produce more than 200 tokens per second. It costs $0.15 for every million input tokens (about 200 Korean won) and $0.6 for every million output tokens (about 800 Korean won). This means the same AI agent can do its work for less than one-third of the previous cost.

DeepSeek, another Chinese AI company, released an upgraded model called "V4.1-Flash-1017" on the 17th. It helps tasks finish faster and gives shorter answers to reduce the cost of each task. DeepSeek said it had improved the model's ability to reason, write code, work as an AI agent, develop games, and create 3D models. The model costs $0.3 per million input tokens and $1.2 per million output tokens. For coding tasks, the average cost for 1 cases fell from $0.5 to less than $0.2 with the new model.

Both MiniMax and DeepSeek designed their new models for the AI agent market. More people are using AI agents to do things such as write code, search for information, shop, and schedule activities. As a result, speed and cost have become important factors when choosing an AI model. A technology industry source said, "Most U.S. and Chinese AI models perform similarly for simple tasks when used as agents. For complex tasks, advanced U.S. AI is preferred, while cost-effective Chinese AI is often used for general, diverse tasks." Chinese AI companies are using this difference to their advantage. They are making their models faster and cheaper so users can get more work done with several AI agents.

U.S. AI companies have also been making progress, releasing more than 10 models over the past month. The Information reported on the 19th (local time) that Anthropic, Google, and OpenAI are speeding up the release of new models. Anthropic is preparing a model with stronger reasoning and intuition that can understand what users mean even when they do not say it directly. Google is developing "Gemini 4 Carbon," an improved version of "Gemini 4 Argon." The earlier model was praised for handling visual tasks well, including video analysis and website testing. OpenAI is expected to release a mid-level model called "Terra." It is expected to perform better while costing less than GPT-6.1 Sol. OpenAI is also testing the safety and security of a next-generation flagship model that uses a design called a "looping transformer."

Industry experts say the "AI development speed control theory" proposed by technology leaders a month ago has become a "vain prayer," meaning it is unlikely to succeed. A technology industry source said, "Only a few companies have established external evaluation processes for their AI, while no real efforts to slow development exist. Amid ongoing U.S.-China competition, advocating for speed control is absurd."