Chinese AI models cost less as Kimi K3 challenges U.S. labs

Chinese AI models cost far less than leading U.S. systems while getting close enough in performance to win business from developers and major companies, according to Fortune. The shift has put Moonshot AI’s Kimi K3, Z.ai’s GLM models and DeepSeek at the center of a new challenge to U.S. AI labs.

Moonshot AI, based in Beijing, released Kimi K3 on July 16 and described it as the largest open-source model yet released. The company said the model can perform near Anthropic’s Fable 5 at a much lower price; Moonshot’s benchmarks put K3 among the top three AI models, while Arena.AI ranked it first among available models, Fortune reported.

The release raised doubts about whether U.S. companies can hold their lead by spending more on computing power. Asian markets sold off after the launch, the Philadelphia Semiconductor Index fell 1.6%, and Nvidia lost almost $600 billion in market value while briefly giving up the world’s most valuable company title to Apple, according to Fortune.

Why are Chinese AI models cheaper?

Chinese labs have been forced to get more output from less advanced hardware because U.S. export controls limited their access to top AI chips beginning in 2022. Analysts cited by Fortune said lower power costs, heavy use of open-source licenses and a willingness to accept thinner margins have also helped Chinese companies cut prices.

Paul Triolo, a partner at DGA–Albright Stonebridge Group, told Fortune that China’s AI ecosystem appears stronger than many observers had assumed. Grace Shao, an AI analyst and author of the AI Proem newsletter, said Chinese labs are careful with resources because they face constraints in computing power, capital and talent.

The price gap is wide. Fortune reported that 1 million output tokens, about 750,000 words, cost $50 on Anthropic’s Fable model. The same amount costs about $0.87 on DeepSeek-V4-Pro, $4.40 on Z.ai’s GLM-5.2 and $15 on Kimi K3, according to figures cited by Fortune from Moonshot AI and OpenRouter.

DeepSeek helped shift views of China’s AI prospects in early 2025 with its V3 and R1 models, which the Hangzhou lab said it trained on a small budget through efficiency gains. In June, Z.ai drew attention with GLM-5.2, a model Fortune said was strong in coding and creative design; its stock had risen more than 1,100% through mid-July before falling 40% over two days after Kimi K3’s release.

Chinese internet companies are also entering the race. Meituan said its LongCat-2.0 model was trained entirely on Chinese-made processors and performed at levels comparable with OpenAI and Anthropic releases from February, according to Fortune.

Which companies are using Chinese AI models?

Chinese models have gained usage on OpenRouter, a marketplace that lets developers access multiple AI models through one interface. Fortune reported that at one point in mid-July, six of the top 10 models on the platform came from Chinese companies including Tencent, Xiaomi, DeepSeek, MiniMax, Moonshot and Z.ai; Chinese models accounted for 57% of tokens used by U.S. firms on OpenRouter in one July week.

Adoption has moved beyond developer testing. Airbnb CEO Brian Chesky said last year that the company used Alibaba’s Qwen for customer service, while Cursor has said Moonshot’s Kimi was the basis for Composer 2, its coding model. Coinbase CEO Brian Armstrong wrote in a June social media post that the company cut its AI spending in half by increasing use of Kimi and Z.ai’s GLM models, Fortune reported.

DoorDash is also assigning some coding work to Chinese models. Chief technology officer Andy Fang said the company sends “lower-level work” to Kimi, producing “better quality [at] cheaper cost,” according to Fortune.

The rise of Chinese models has drawn scrutiny in Washington. Congress is examining the use of Chinese AI models by U.S. companies including Airbnb and Cursor, Fortune reported; Airbnb said it used only “a limited number” of Chinese open-source models through approved U.S.-based service providers, while Cursor did not respond to Fortune’s request for comment.

Chinese developers are presenting open access as an advantage as U.S. policy becomes less predictable for foreign users. Z.ai said after announcing GLM-5.2 that frontier intelligence should not be controlled by a few people or withdrawn under changing rules, Fortune reported.

This story draws on original reporting from Fortune.