9.4×10²⁶
FLOP: Epoch's estimate for GPT-6 Astra (Sep 2026), the largest training run in its database
Largest run to date (log scale)
20102026
What the chart shows
Training compute is the total number of calculations used to train a model. Four notable models in Epoch’s database exceed 10²⁶ FLOP: GPT-6 Astra (an estimated 9.4×10²⁶, September 2026), Grok 4 (5×10²⁶), GPT-4.5 (3.8×10²⁶) and Grok 3 (3.5×10²⁶). All come from US companies.
The median notable model of 2025 used about 3.9×10²⁴ FLOP, roughly 550,000 times the 2015 median of 7×10¹⁸.
Key numbers
| Measure | Value |
|---|---|
| GPT-6 Astra (Sep 2026) | 9.4×10²⁶ |
| Grok 4 (Jul 2025) | 5×10²⁶ |
| GPT-4.5 (Feb 2025) | 3.8×10²⁶ |
| Median, 2025 (40 models) | 3.9×10²⁴ |
| Median, 2015 (11 models) | 7×10¹⁸ |
How to read it
- Compute for most frontier models is estimated by Epoch from hardware, training time and other disclosures, not reported by developers.
- The vertical axis is logarithmic: each gridline is ten times the one below.
- 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 1.1.13. Includes models published up to 30 September 2026.
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