Chinese AI Models Gain Ground in Global Coding Market
Chinese artificial intelligence (AI) models are rapidly eroding the market share of U.S. models such as Anthropic's Claude and Google's Gemini in the global coding market. Beyond their price competitiveness, open-source models with performance comparable to leading U.S. models are emerging, expanding their influence. Notably, their expansion in emerging markets like the Middle East and South America is significant.
According to the information technology (IT) industry on the 7th, the global share of Claude and Gemini in the programming sector has steadily decreased, while the share of Chinese AI has increased significantly.
According to OpenRouter, Anthropic's Claude SONET 4 share in the programming area as of Aug. 11 fell 15.7 percentage points compared to July 21. The Gemini 2.5 Pro and Flash also decreased by 3.6 percentage points and 4.4 percentage points, respectively.
On the other hand, Alibaba's Qwen 3 coder grew 16.4% over the same period, accounting for 21.5% of the market. Qwen's growth was particularly notable. While DeepSeek is slowing down, it has ranked first among Chinese models in terms of performance. Alibaba's frontier model "Qwen3-235B-A22B-Thinking-2507" scored 64 points, ahead of DeepSeek's latest model V3.1 (60 points), according to Artistic Analytics indicators.
Chinese start-ups are also catching up. Z, released in July, according to the same survey by OpenRouter.AI's GLM4.5 and Moonshot AI's Kimi-K2 had a 6.1% and 3.2% share as of Aug. 11, respectively.
Among them, Kimi-K2 attracted significant attention, evaluated as bringing another "deep moment." According to Moonshot AI, in the programming-specific benchmark "SWE-Bench Verified," Kimi-K2 ranked second behind Anthropic Claude Opus 4 (72.5%) with 65.8% accuracy. Z.AI's GLM4.5 also outperformed OpenAI's o3 model on its own benchmark.
An industry official stated, "Now, not only China's DeepSeek but also most Chinese open-source models are competitive to compete with U.S. big tech models in performance beyond cost-effectiveness."
There are many performances, but the biggest reason why the Chinese model stands out in the global market is also price competitiveness. The Qwen3 coder costs $1 per 1 million tokens of input and $5 per 1 million tokens of output, which is 15 times cheaper than the Claude Opus 4 (input $15, output $75).
The startup model is more unconventional. Z.AI's GLM4.5 is the lowest among Chinese models, at $0.6 per million tokens of input and $2.2 output. MoonshotAI's Kimi-K2 is also very affordable, with $0.6 input and $2.5 output.
This price competitiveness is particularly strong in emerging countries such as the Middle East and South America than in the United States or Northeast Asia. Qwen and Z. Chinese models such as AI are showing rapid growth in the coding market in emerging countries, and it is analyzed that price competitiveness is effective in the background.
Similar web data showed that Qwen models, excluding China, accounted for 27.5% of traffic in Iraq, 19.1% in Brazil and 12.1% in Turkiye.
Z.AI also has offices in the Middle East and Africa to supply AI solutions to local governments and state-owned companies. OpenAI has recently pointed to the company as a check target, saying it is winning government AI supply contracts in emerging countries with the support of the Chinese Communist Party.
As OpenAI pointed out, the rise of China's open-source model is not just a corporate-level result. Since the "deep shock" earlier this year, China has established an open-source strategy at the national level and provided full support to related companies.
This year, Chinese companies launched a series of frontier models and the rise in the global share of Chinese models is also attributed to the government's support. The recent fact that Alibaba Qwen-based derivatives alone exceeded 100,000 shows the pace of expansion of China's open-source ecosystem.
As such, the Chinese model has emerged rapidly this year and is competing with the U.S. model in the global coding market, while the Korean model's presence in this market is still insignificant.
Although LG AI Research Institute's recently released 'Exemployee 4.0' was evaluated as being at the top of the global rankings in coding performance, it is far from the actual market share. The reality is that many domestic companies also use overseas models.
Industry sources point out that it is urgent to strengthen capabilities in the coding sector, one of the key areas of AI model competitiveness, as discussions on the development of "Sovereign AI" have recently been in full swing in Korea.