US AI Models Lead China by 7 Months: No Chinese AI Has Yet Matched OpenAI’s o3

Since 2023, a measurable gap has emerged between the most capable AI models from the US and China. Chinese models reach the capabilities of their US counterparts with an average delay of seven months, as an analysis of the Epoch Capabilities Index (ECI) shows. The range spans from a minimum of four to a maximum of 14 months – a gap that persists throughout development since early 2023.

Every model that has reached the frontier of AI capabilities since 2023 originates from the United States. Models like GPT-4, o1, and Gemini 3 Pro Preview set the benchmarks against which Chinese developments such as DeepSeek-R1, Qwen2.5-72B, or Yi-34B must measure themselves. The ECI data document this dominance comprehensively over a three-year period.

Open Weight versus proprietary

The gap between US and Chinese models closely resembles the general divide between proprietary and open-weight models. This parallel comes as little surprise: nearly all leading Chinese models are released as open-weight, while US frontier models remain closed. DeepSeek-V2, Qwen-14B, and Baichuan2-13B exemplify China’s open-weight strategy, which creates transparency but apparently cannot match the performance of closed US systems.

Concrete milestones illustrate the lag: the first Chinese model to surpass GPT-4 reached this level in May 2024 – 14 months after OpenAI’s flagship release in March 2023. OpenAI’s o3 model, which launched in April 2025, has yet to be matched or exceeded by any Chinese model. The gap persists.

Measurement methodology

Epoch AI calculates the gap through precise comparison: for each day, researchers determine the ECI score of the best Chinese model and then establish how long ago the leading US model achieved a comparable or lower score. Models are considered performance-equivalent when their scores differ by a maximum of one ECI point. This method enables day-precise quantification of the gap.

Luke Emberson’s analysis deliberately excludes the initial models – LLaMA-65B for the US and Baichuan1-7B for China – since these likely did not represent the actual frontier when ECI data collection began in January 2023. The data encompass all relevant releases through early 2026 and paint a consistent picture: the US leads, China follows with a seven-month delay.