TL;DR
Epoch’s benchmarking shows that every model at the measured AI frontier since 2023 was developed in the United States. Chinese models lagged that frontier by an average of seven months, with gaps ranging from four to 14 months.
What happened
Epoch Analytics used its Epoch Capabilities Index (ECI) to map relative model capabilities since 2023 and found that every model at the frontier, as defined by that index, originated in the United States. Over the period analyzed, Chinese models consistently trailed the US frontier: the average lag was seven months, with the shortest gap at four months and the longest at 14 months. The dataset and visualization note that this pattern mirrors a broader difference between proprietary (closed-weight) and open-weight models — most leading Chinese models are open-weight, while the frontier US models remain closed. Epoch published the analysis on January 2, 2026 (author: Luke Emberson) and makes the work available under a Creative Commons BY license.
Why it matters
- A sustained lead by US-developed frontier models indicates a persistent capability gap as measured by Epoch’s ECI.
- The pattern aligns with distinctions between closed/proprietary and open-weight model development approaches, which may influence access and distribution.
- Benchmarking that consistently places frontier models in one country can shape discussions around research priority, policy, and international competitiveness.
- The result connects to broader data insights from Epoch showing US outperformance by non-US models and hardware distribution differences.
Key facts
- Metric used: Epoch Capabilities Index (ECI) to compare model capabilities.
- Since 2023, every model at the frontier measured by the ECI was developed in the United States.
- Average lag for Chinese models versus the US frontier: seven months.
- Observed lag range for Chinese models: minimum four months, maximum 14 months.
- Epoch notes the capability gap resembles the broader gap between proprietary (closed-weight) and open-weight models.
- Nearly all leading Chinese models referenced in the analysis are open-weight; frontier US models are described as closed.
- Examples of models appearing in the Epoch visualization include Gemini 3 Pro Preview, GPT-5, o1, o1-mini, GPT-4 (Mar 2023), DeepSeek-R1, Qwen2.5-72B, DeepSeek-V2 (MoE-236B), Yi-34B (chat), Qwen-14B, Baichuan2-13B, and Qwen3 Max.
- Analysis author: Luke Emberson; publication date: January 2, 2026.
- Epoch’s work is released under a Creative Commons BY license and is free to use with attribution.
What to watch next
- Whether future Epoch ECI updates show Chinese models closing the observed gap (not confirmed in the source).
- Any shifts in the balance between open-weight and closed-weight leading models across countries (not confirmed in the source).
- New model releases from US or Chinese developers that alter the timeline of frontier status as tracked by the ECI (not confirmed in the source).
Quick glossary
- Epoch Capabilities Index (ECI): A comparative scoring method used by Epoch to estimate and visualize relative capabilities of AI models over time.
- Open-weight model: A model whose trained weights are publicly released, permitting inspection, reuse, and modification by others.
- Closed-weight (proprietary) model: A model whose trained weights are not publicly shared, typically controlled by the developing organization.
- Frontier model: In this context, a model that ranks at the leading edge of capabilities according to the chosen benchmark or index.
Reader FAQ
What is the main finding of Epoch’s analysis?
Epoch reports that since 2023 every model at the measured frontier was developed in the US, and Chinese models trailed by an average of seven months.
How large was the gap between Chinese and US models?
The analysis found an average lag of seven months, with a minimum observed gap of four months and a maximum of 14 months.
Does the report say why the gap exists?
Not confirmed in the source.
Are the findings publicly available to use?
Yes. Epoch’s work is available under a Creative Commons BY license and can be reused with attribution.
Latest Publications & Commentary Papers & Reports Newsletter Podcast Data & Resources Datasets Overview Benchmarking Models Frontier Data Centers ML Hardware AI Companies GPU Clusters Polling Resources AI Trends &…
Sources
- Chinese AI models have lagged the US frontier by 7 months on average since 2023
- Chinese AI models have lagged the US frontier by 7 …
- Nano Banana 2 Flash 🍌, Grok Business 💼, Plaud AI …
- Top 10 China AI Stories in 2025: A Year-End Review
Related posts
- Nvidia reportedly asks Chinese buyers to pay upfront for H200 GPUs
- At CES 2026, TV makers pile on intrusive AI features to sell flat market
- AI-altered Images Are Falsely Identifying the Agent Who Shot Renee Good