38×
GPT-6 Astra's estimated compute (9.4×10²⁶ FLOP) relative to China's largest, Doubao-pro (2.5×10²⁵)
What the chart shows
The most compute-intensive US model in Epoch’s database, GPT-6 Astra (September 2026), used an estimated 9.4×10²⁶ FLOP. China’s largest, ByteDance’s Doubao-pro (October 2024), used about 2.5×10²⁵, so the US run used roughly 38 times as much.
Chinese developers release many notable models (23 with compute estimates in 2025, against 15 from the US), but at the top end the US models are an order of magnitude larger.
Key numbers
| Measure | Value |
|---|---|
| Largest US: GPT-6 Astra (Sep 2026) | 9.4×10²⁶ |
| Largest China: Doubao-pro (Oct 2024) | 2.5×10²⁵ |
| Largest China in 2026: Kimi K3 (Jul 2026) | 2×10²⁵ |
How to read it
- Compute is mostly estimated by Epoch from public information.
- A model from organisations in both countries appears in both series.
- Includes models published up to 30 September 2026.
Source: Epoch AI, Data on AI models (CC BY 4.0), accessed 19 October 2026. Analysis and chart by AI Adoption, following the method of the AI Index Report 2026 (Stanford HAI), figure 1.1.14. Shows all notable models from each country with a compute estimate, so it includes more models than the report's selection.
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Discussion