Real user prompt benchmark from WildChat — evaluates instruction-following quality on diverse, authentic user prompts scored by GPT-4-turbo on a 1-10 scale, rescaled to a WB Score.
Uses genuine user prompts (not synthetic), capturing the true distribution of tasks people ask LLMs to do. Length-penalized scoring variant available.
GPT-4-turbo as judge introduces systematic biases. The rescaling formula (Y-5)*2 can produce negative scores for weak models.
Higher is better. WB Score (typically -20 to 60 range).
GPT-4-turbo as judge introduces systematic biases.
| # | Model | Score |
|---|---|---|
| 1 | 62.6 | |
| 2 | 60.0 | |
| 3 | 58.9 | |
| 4 | 58.8 | |
| 5 | 58.8 | |
| 6 | 58.4 | |
| 7 | 58.1 | |
| 8 | 57.6 | |
| 9 | Gemini 2.5 Flash-Lite Preview (Sep '25) (Non-reasoning) source benchpress Evidence observed Apr 13, 2026 | 57.6 |
| 10 | 57.4 | |
| 11 | 57.3 | |
| 12 | 57.3 | |
| 13 | 57.2 | |
| 14 | 57.1 | |
| 15 | 57.1 | |
| 16 | 56.5 | |
| 17 | 56.2 | |
| 18 | 56.1 | |
| 19 | 56.0 | |
| 20 | 56.0 | |
| 21 | 56.0 | |
| 22 | 55.9 | |
| 23 | 55.8 | |
| 24 | 55.8 | |
| 25 | 55.7 | |
| 26 | 55.1 | |
| 27 | 55.1 | |
| 28 | 55.0 | |
| 29 | 54.9 | |
| 30 | 54.7 | |
| 31 | 54.6 | |
| 32 | 54.5 | |
| 33 | 54.4 | |
| 34 | 54.4 | |
| 35 | 54.3 | |
| 36 | 54.3 | |
| 37 | 54.1 | |
| 38 | 53.9 | |
| 39 | 53.8 | |
| 40 | 53.8 | |
| 41 | 53.8 | |
| 42 | 53.7 | |
| 43 | 53.7 | |
| 44 | 53.7 | |
| 45 | 53.6 | |
| 46 | 53.5 | |
| 47 | 53.3 | |
| 48 | 53.1 | |
| 49 | 53.1 | |
| 50 | 53.0 |