Chinese AI companies released 16 models in September as Amodei urged a slowdown

Chinese AI companies released 16 new models in September, the same month Anthropic Chief Executive Dario Amodei called for slower advances in AI capabilities. A Nikkei Asia tally counted releases from 10 leading developers, including DeepSeek, Alibaba (BABA), Tencent (TCEHY) and Xiaomi.

The count includes newly developed text, image and audio models, but excludes adaptations of other companies' systems. Nikkei didn't publish the full list of 16.

Most of the new models are built for lower-cost use

DeepSeek released V4.1-Flash on Sept. 10 and cut its API prices the same day. DeepSeek describes the model as a 552-billion-parameter mixture-of-experts design with "just 8B active parameters for input, 16B for output." DeepSeek is also close to raising at least 80 billion yuan in a round backed by Tencent.

Other lightweight releases in September included Z.ai's GLM-5.3-FlashX and MiniMax's M3.1-Flash-Preview, along with similar models from Alibaba and Xiaomi. Demand for cheaper models is growing as AI agents catch on, because agents use large numbers of tokens, the units of data that AI providers process and bill by.

Leading US models still score higher on benchmarks, according to the research firm Artificial Analysis. Nikkei put the cost of Anthropic's most advanced models at roughly 20 to 50 times that of newer DeepSeek and Xiaomi models, without specifying whether it compared per-token prices or task-completion costs. It said the price gap is widening even as the performance gap narrows.

Chinese companies are building bigger models too

At its annual cloud conference in September, Alibaba announced plans to develop a version of its Qwen model with 5 trillion to 10 trillion parameters. Its latest flagship, unveiled in August, has 2.4 trillion.

iFlytek and state-owned China Telecom released models in September that use Chinese-made chips for both training and running them. US export controls limit China's access to advanced AI chips.

The pace has picked up over the year. Eight major Chinese developers released 20 large language models from July through September, up from 17 in the second quarter and two in the first, according to the benchmarking site BenchLM. US developers released 59 from January through September, about 50% more than China.

Amodei's essay warned China could pull ahead

Amodei called for slowing AI progress in an essay published Sept. 12. "We must slow the pace at which we improve the capabilities of AI models," he wrote. OpenAI Chief Executive Sam Altman backed the call the same day.

Amodei said pacing would allow more time for safety work and outside evaluation while training continued. His proposal concerns how quickly capabilities improve. Nikkei's tally counts model releases.

The essay tied any slowdown to the US lead over China. It said, "Pacing within democracies will be limited by the lead that US companies have over authoritarian regimes, chiefly the Chinese Communist Party." Amodei added, "If we slow down by more than this amount, then (unpaced) CCP-associated projects will pull ahead, creating significant national security risk."

To protect that lead, Amodei called for no sales of powerful AI chips or chipmaking equipment to China. He also urged a crackdown on "unauthorized distillation by companies in authoritarian countries," a technique that trains a model on a stronger model's outputs. In February, Anthropic accused DeepSeek, Moonshot AI and MiniMax of distilling Claude.

"China, which places importance on using AI to advance industry and scientific research, will not slow development," Ichiro Satoh, a professor at Japan's National Institute of Informatics, told Nikkei.

DeepSeek published a paper in September on techniques for safely training AI agents, including methods designed to limit abnormal behavior. The paper said no single mechanism can prevent that behavior and that more technical work is needed. DeepSeek Chief Executive Liang Wenfeng is listed as an author.