Total floating-point operations (FLOP) used during model training, as estimated by Epoch AI from public disclosures, hardware counts, and training duration.
Directly measures training investment. Sourced from Epoch AI's systematic research covering 1,600+ frontier models.
Estimates vary by methodology. Does not account for inference-time compute or post-training optimization.
Lower is better for efficiency at similar capability. Values are total training FLOP (for example 1e25), so interpret alongside quality benchmarks.
Estimates vary by methodology.
| # | Model | Score |
|---|---|---|
| 1 | Switchpoint Router | 8.22e+22 |
| 2 | Writer: Palmyra X5 | 2.5272e+24 |
| 3 | OpenAI: GPT-3.5 Turbo Instruct | 2.578e+24 |
| 4 | Claude 2.0 | 3.866e+24 |
| 5 | PALM-2 | 7.34e+24 |
| 6 | GPT-4 | 2.1e+25 |
| 7 | Claude 3.5 Sonnet (Oct '24) | 2.700000000000001e+25 |
| 8 | Grok 2 (Dec '24) | 2.96e+25 |
| 9 | Meta: Llama 3.1 405B Instruct | 3.8e+25 |
| 10 | Gemini 1.0 Ultra | 5.0000000001e+25 |
| 11 | Grok 3 | 3.5e+26 |
| 12 | GPT-4.5 (Preview) | 3.8e+26 |
| 13 | Grok 4 | 5.0000000000001e+26 |