The development cycle of AI models in China and the United States has been shortened to 1/3, with an average of 44 days.

Nikkei has counted the update cycles of high-performance models from 5 U.S. companies and 4 Chinese companies. The average interval between January 2023 and March 2026 was 125 days, while it shrank sharply to 44 days from April to September 2026. The concerns raised by the "self-evolution" where AI itself participates in AI development are also intensifying day by day...

The artificial intelligence (AI) development in China and the United States is accelerating. Since April, the cycle of new model releases has been shortened to an average of 44 days (about 1.5 months), only one-third of the previous level. The "self-evolution" where AI itself participates in AI development is taking place, which has become the trigger for the AI threat theory.

OpenAI and Anthropic released their latest models in September

Nikkei surveyed 5 leading U.S. companies in AI model development including Anthropic and OpenAI, as well as 4 Chinese companies including Alibaba Group and Moonshot AI, and counted the cycles for each company to update their high-performance models.

The average interval between January 2023 and March 2026 was 125 days, while it dropped drastically to 44 days from April to September 2026.

Entering September, U.S.-based OpenAI and Anthropic successively released their latest models. Since July, Meta has been upgrading its cutting-edge model "Muse Spark" every month. Google in the U.S. launched the updated Gemini version that outperforms the previous one on September 2, taking only three weeks. And SpaceX's Grok launched new models for three consecutive months up to August.

For Chinese enterprises, DeepSeek has carried out model updates every month since July. After August, Alibaba and Z.AI (Beijing Zhipu Huazhang Technology), which launched the "GLM" series, also successively released new models.

One of the reasons for the accelerated development speed is that AI itself is undertaking the development of AI models. Nowadays, AI is capable of handling difficult tasks such as experiment monitoring and result analysis.

Data released by Anthropic on September 17 shows that as of August, 26% of its development work was led by AI, and more than 90% of the work involved AI participation. Back in February this year, the proportion of AI-led development was almost zero. Anthropic pointed out, "It is possible to form a virtuous cycle where AI independently develops next-generation AI models with higher performance."

At OpenAI, the operating time of AI is 3.1 times that of humans

At OpenAI, the operating time of AI agents in August reached 3.1 times the working hours of human researchers.

Converted by monetary value, the average regular researcher at the company uses AI agents worth about 600 U.S. dollars per day. The top 10% of researchers assign more than 7,000 U.S. dollars worth of workload to AI every day.

As of early June, human researchers had longer working hours, but the role of AI has increased rapidly since then.

With the widespread application of AI, the amount of code written through programming at OpenAI in August reached 7 times the average level in 2025. The amount of code actually applied to products and other scenarios by Anthropic from April to June also reached 8 times the average level from 2021 to 2025. The increase in code volume helps improve the performance of AI models.

The speed of AI performance improvement is also accelerating. Data from the UK government agency Artificial Intelligence Safety Institute (AISI) shows that the time required for the cyberattack capability of AI models to double was 4.7 months in February 2026, which is shorter than the 8 months recorded in November 2025. It is said that this time has been further shortened currently.

According to the AI performance evaluation index created by U.S. research firm Artificial Analysis combining multiple tests, the capabilities of AI models have improved significantly after entering 2026.

At present, the models of the U.S. camp are upgraded more frequently. The top 5 U.S. companies released 20 AI models from July to September, doubling the number from April to June.

Models are segmented with features such as low price and speed priority

While pursuing performance improvement, the trend of segmenting models for use cases such as low price, speed priority and focus on cybersecurity is also expanding. As Chinese enterprises launch "open-source" high-performance AI models that everyone can use, U.S. enterprises are also seeking to maintain competitive advantages by expanding their product lines.

Concerns that AI's self-evolution will lead to situations out of human control are also intensifying. In the United States, centered on Dario Amodei, Chief Executive Officer (CEO) of Anthropic, calls for slowing down AI model development are rising continuously.

The report released by Anthropic on September 17 points out: "A third-party institution should assess the role played by AI in the AI model development process, and whether humans can properly supervise the AI involved in development."

At OpenAI, there has been an accident where AI misjudged the tasks to be completed and "escaped" from the development environment.

Against the backdrop of continuously improving AI performance, if governance and safety management issues are ignored, more serious accidents may occur in the future. Whether China and the United States, which are setting off a development race, can reach a consensus on establishing a safety management system will become an increasingly important issue.

This article is from the WeChat Official Account "Nikkei Chinese Net" (ID: rijingzhongwenwang), authored by Ai Mi Xiao and Yito Kishima, and published with authorization from 36Kr.