Artificial intelligence development in the US and China is picking up pace as AI assumes an increasingly large role in its own development, fueling concerns about the risk posed by advanced systems.
Nikkei looked at five leading US players, including Anthropic and OpenAI, and four Chinese companies, including Alibaba and Moonshot AI, to examine the time required to release upgraded versions of high-performance models.
Between January 2023 and March 2026, the average interval between releases was 125 days. But from April to September of 2026, this shrank to 44 days.
In September, OpenAI and Anthropic released their latest models in quick succession. Meta Platforms has been introducing updates to its flagship model, Muse Spark, every month since July. Google released a new version of Gemini on September 2, just three weeks after releasing its previous version. SpaceXAI’s Grok also launched new models in the three months through August.
Chinese developers are moving at a similar clip. DeepSeek has updated models monthly since July. Since August, Alibaba and Z.ai, developer of the GLM series, have also released a string of new models.
One reason for the faster pace is that AI itself is now handling AI research and development. AI systems are increasingly capable of performing such sophisticated tasks as monitoring experiments and analyzing results.
In a September 17 blog post, Anthropic said that as of August its Claude AI “leads” 26% of its R&D efforts and was involved in more than 90% of all R&D activities. In February, the share of AI-led research was practically zero.
“However, models accelerating their own development could make it more challenging for humans to understand or control these systems,” Anthropic wrote in the blog post.
At OpenAI, the number of hours worked by AI agents in August was more than triple that of human researchers. Human researchers were contributing more work hours than AI systems as recently as June.
As AI adoption has picked up, the volume of programming code written at OpenAI in August rose to seven times the 2025 annual average. At Anthropic, the amount of code actually incorporated into products between April and June was eight times the average level from 2021 through 2025.
More code generally translates into improved model performance.
The pace of performance improvements is also increasing. The time required for an AI model’s cyberattack capabilities to double was 4.7 months as of February 2026, according to the UK’s AI Security Institute, down from eight months in November 2025. The time needed has since further shortened, according to the institute.
An AI performance index that combines multiple benchmarks compiled by US research firm Artificial Analysis also shows gains in capability. Since 2025, such companies as DeepSeek have risen to prominence with models excelling at deriving answers through logical inference. Heading into 2026, these reasoning capabilities increasingly became the foundation for AI agents capable of autonomously performing a wide variety of tasks.
China’s AI models are generally estimated to lag those in the US by four to six months. But since June, Chinese developers have begun releasing highly capable models with advanced agent functionality, and AI’s role in AI development is expected to continue growing.
Since July, larger-scale models have appeared more frequently, indicating increased investment in computing resources.
Facing pressure from Chinese competitors, US companies have accelerated model releases. The top five US developers announced 20 AI models between July and September, twice the number released from April through June.
Rather than focusing solely on performance, developers are increasingly tailoring models for specific needs, including low-cost operations, faster response times and enhanced cybersecurity. As Chinese companies provide high-performance open models available to everyone, American companies are broadening their lineups to preserve their competitive advantage.
Concerns are growing that self-evolving AI could eventually become difficult or impossible for humans to control. In the US, calls to slow the pace of AI development, including those by Anthropic CEO Dario Amodei, are growing louder.
Anthropic argued in its September 17 blog post that independent organizations should verify both the role AI plays in AI R&D and whether humans are adequately supervising AI systems involved in this work.
OpenAI has also reported incidents where AI systems ignored given instructions and escaped their isolated development environments.
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.
