Enterprise AI in Numbers · Capability and cost
Capability gap between the best US and the best Chinese AI model
11.0 points
11.0 points (Claude Opus 5.5 vs Kimi K3) as of September 22, 2026, according to Epoch AI, Epoch Capabilities Index, up from 10.9 points on September 3, 2026. The ECI of the most capable model from a US organization minus that of the most capable model from a Chinese organization.
What it means for leaders
Chinese models, often open-weight and cheap, trail the US frontier by months rather than years. For sovereignty, export control and data residency decisions, assume capable alternatives exist on both sides.
What would change the reading: The gap narrowing after a major Chinese release, or widening after a US one.
The trend
Every reading
| Date | Value | Change | Detail | Source |
|---|---|---|---|---|
| 11.0 points | +0.1 points | Claude Opus 5.5 vs Kimi K3 | Source | |
| 10.9 points | +2.3 points | GPT-6 Astra vs Kimi K3 | Source | |
| 8.6 points | +2.5 points | Claude Fable 5.1 vs Kimi K3 | Source | |
| 6.1 points | +0.5 points | Claude Opus 5 vs Kimi K3 | Source | |
| 5.6 points | −3.8 points | Claude Fable 5 vs Kimi K3 | Source | |
| 9.4 points | +0.5 points | Claude Fable 5 vs Qwen3.7-Max | Source | |
| 8.9 points | −3.1 points | GPT-5.5 Pro vs Qwen3.7-Max | Source | |
| 12.0 points | +3.4 points | GPT-5.5 Pro vs Kimi K2.6 | Source | |
| 8.6 points | −1.2 points | GPT-5.4 Pro vs Kimi K2.6 | Source | |
| 9.8 points | −1.7 points | GPT-5.4 Pro vs GLM-5.1 | Source | |
| 11.5 points | +2.5 points | GPT-5.4 Pro vs Kimi K2.5 | Source | |
| 9.0 points | +0.7 points | GPT-5.3 Codex vs Kimi K2.5 | Source | |
| 8.3 points | −1.9 points | GPT-5.2 Pro vs Kimi K2.5 | Source | |
| 10.2 points | +3.2 points | GPT-5.2 Pro vs DeepSeek-V3.2 | Source | |
| 7.0 points | −0.3 points | Gemini 3 Pro vs DeepSeek-V3.2 | Source | |
| 7.3 points | +2.5 points | Gemini 3 Pro vs Kimi K2 Thinking | Source | |
| 4.8 points | −0.9 points | GPT-5 Pro vs Kimi K2 Thinking | Source | |
| 5.7 points | +0.6 points | GPT-5 Pro vs DeepSeek-V3.2-Exp | Source | |
| 5.1 points | −1.1 points | GPT-5 vs DeepSeek-V3.2-Exp | Source | |
| 6.2 points | +2.6 points | GPT-5 vs Qwen3-235B-A22B-Thinking (Jul 2025) | Source | |
| 3.6 points | −2.6 points | o3-pro vs Qwen3-235B-A22B-Thinking (Jul 2025) | Source | |
| 6.2 points | +0.6 points | o3-pro vs DeepSeek-R1 (May 2025) | Source | |
| 5.6 points | −2.0 points | o3 vs DeepSeek-R1 (May 2025) | Source | |
| 7.6 points | −0.3 points | o3 vs Qwen3-235B-A22B | Source | |
| 7.9 points | +2.8 points | o3 vs DeepSeek-R1 | Source | |
| 5.1 points | +2.2 points | Gemini 2.5 Pro (Mar 2025) vs DeepSeek-R1 | Source | |
| 2.9 points | −6.5 points | o1 vs DeepSeek-R1 | Source | |
| 9.4 points | −3.4 points | o1 vs DeepSeek-V3 | Source | |
| 12.8 points | +6.1 points | o1 vs Qwen2.5-72B | Source | |
| 6.7 points | −3.7 points | o1-mini vs Qwen2.5-72B | Source | |
| 10.4 points | +5.7 points | o1-mini vs Qwen2-72B | Source | |
| 4.7 points | +1.0 points | Claude 3.5 Sonnet vs Qwen2-72B | Source | |
| 3.7 points | −0.5 points | GPT-4o (May 2024) vs Qwen2-72B | Source | |
| 4.2 points | +1.7 points | GPT-4o (May 2024) vs DeepSeek-V2 (MoE-236B, May 2024) | Source | |
| 2.5 points | −7.4 points | GPT-4 Turbo (Apr 2024) vs DeepSeek-V2 (MoE-236B, May 2024) | Source | |
| 9.9 points | +0.3 points | GPT-4 Turbo (Apr 2024) vs Yi-34B | Source | |
| 9.6 points | +0.5 points | Claude 3 Opus vs Yi-34B | Source | |
| 9.1 points | +0.7 points | GPT-4 Turbo (Nov 2023) vs Yi-34B | Source | |
| 8.4 points | −4.5 points | GPT-4 (Mar 2023) vs Yi-34B | Source | |
| 12.9 points | −9.9 points | GPT-4 (Mar 2023) vs Qwen-14B | Source | |
| 22.8 points | −12.9 points | GPT-4 (Mar 2023) vs Baichuan2-13B | Source | |
| 35.7 points | — | GPT-4 (Mar 2023) vs Baichuan1-7B | Source |
Definition and method
- What it measures
- The ECI of the most capable model from a US organization minus that of the most capable model from a Chinese organization.
- Population
- Models scored on the Epoch Capabilities Index (ECI), by the developer's country
- Updated
- When new models are scored
- Source
- Epoch AI, Epoch Capabilities Index · CC BY 4.0
Caveats
- ECI points are a composite across many benchmarks; a gap of a few points can hide larger gaps on specific tasks.
- Most leading Chinese models are open-weight, so the gap also reflects the open versus closed gap.
Cite this
Enterprise AI in Numbers, "Capability gap between the best US and the best Chinese AI model", https://lucianoayres.com/ai-brief/data/us-china-frontier-gap/, retrieved October 25, 2026. Data: Epoch AI, Epoch Capabilities Index.