AI Models
•
July 13, 2026
•
5 min read
Why People Use Chinese AI Models
By Mohid Mirza, Co-Founder of AcceleratedLogic AI
Mohid Mirza
Co-Founder of AcceleratedLogic AI
Right now, there’s no debate that American AI models are the absolute best on the market. They dominate the leaderboards on almost every benchmark, such as DeepSWE or ARC-AGI 3, and also are a lot better at answering questions and writing in general. However, Chinese AI models are getting closer and closer to the frontier. It seems like every week, a new Chinese AI model comes out, whether it be the new Qwen Model, or a new release by DeepSeek or Z.ai. However, Chinese models aren’t the very best, so why are people still using them?
On January 20, 2025, DeepSeek released R1, a model that competed directly with OpenAI’s o1 reasoning AI. It was open-source, and 95% cheaper than o1. This caused a massive panic in the stock market, making Nvidia’s stock drop 18% in a single day. DeepSeek was just the beginning of a wider trend: Chinese AI models could compete with American ones.
DeepSeek was followed up by the release of better and better AI models, such as Alibaba’s Qwen 2.5 and Moonshot Kimi K2. Today, there are a handful of Chinese AI companies other than the ones that I mentioned, consisting of but not limited to: MiniMax, LongCat, Xiaomi, Kuaishou, Tencent, ByteDance, StepFun, and Baidu.
Today, the question is not if Chinese AI models are better than American ones, but how much less powerful they are. The answer is not much. In the Artificial Analysis Intelligence Index, which measures how smart an AI model is, the best Chinese AI model, GLM 5.2, scores 51, while Claude Fable 5 scores 60. This gap is incredibly small, when you factor in the price. GLM 5.2 costs $1.40/M Input Tokens to $4.40/M Output Tokens, while Claude Fable 5 costs $10/M Input Tokens and $50/M Output Tokens. However, when you compare GLM 5.2 to an almost equivalent model, like Grok 4.5, then American models are a lot more competitive, because Grok 4.5 costs $2/M Input Tokens and $6/M Output tokens. A more accurate way to judge how expensive a LLM will be, is through using the Artificial Analysis Cost per Intelligence Task Index. GLM 5.2 costs $0.38 per task, while Grok 4.5 costs $0.31 per task. While, this might seem like American models are slightly cheaper, but this is just a piece of the puzzle.
Just take a look at the Artificial Analysis Intelligence Leaderboard, most of the middle section is taken by Chinese AI models.
Artificial Analysis Intelligence Leaderboard
Higher is smarter (Score range 0-100). Red boxes highlight the Chinese frontier AI models.
60
Ant
59
Ope
58
Ope
56
Ope
56
Ant
55
Ope
55
Ope
54
xAI
53
Ant
51
Ope
51
Zhi
51
Met
50
Goo
46
Goo
46
Ali
44
Min
44
Dee
44
Moo
42
MiM
40
Dee
38
Nvi
38
xAI
34
Ali
30
Mis
30
Ant
29
Goo
24
Ope
17
K2
14
Ups
And look at the leaderboard for cost efficiency, the most expensive Chinese model, Qwen 3.7 Max, which is already an outlier, doesn’t even come close to frontier American models
Artificial Analysis Cost per Intelligence Task Index
Lower is cheaper (Cost in USD per benchmark task). Red boxes highlight Chinese cost-leaders.
$0.020
Dee
$0.030
MiM
$0.040
Dee
$0.060
Ope
$0.12
Min
$0.14
xAI
$0.23
Nvi
$0.26
Met
$0.27
Ant
$0.29
Goo
$0.31
xAI
$0.33
Ali
$0.38
Zhi
$0.55
Ope
$0.59
Goo
$0.79
Ali
$0.99
Ope
$1.04
Ope
$1.20
Mis
$1.53
Ant
$1.80
Ant
$2.75
Ant
Also, Chinese AI models are usually open-source, excluding Qwen 3.7 Max, which allows for people to not only run them on their local devices, but also fine-tune them to be more specialized. Being open-source not only makes models pretty much free to run, barring the price of hardware and the cost of electricity, but also makes it safer for enterprise use, as no data will leave the device that the AI is running on.
American AI labs need to start focusing on efficiency, and releasing open-source models. There are real benefits of having efficient models, as they will run faster and cheaper, and most companies are trying to make their models more efficient. However, there really isn’t a push to have open-source AI models, especially from giants such as Anthropic and OpenAI. The former has never had an open-source AI model, and the latter’s was released on August 25, 2025, which, in the world of AI, is equal to decades ago. Gemini has released the Gemma 4 lineup of AI models somewhat recently, but its best model, Gemma 4 31b, is nowhere near as good as frontier models, unless you use AcceleratedLogic’s multi-agent orchestration platforms.