RealToxicityPrompts: leaderboard
Toxic degeneration benchmark with about 100K naturally occurring web prompts, measuring whether model continuations become toxic under realistic prompt prefixes.
Metric: Toxic fraction. Source: github.com. Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Palmyra X (43B) | 0.01 | #1317 |
| 2 | T5 (11B) | 0.01 | #1652 |
| 3 | text-davinci-003 | 0.04 | #1146 |
| 4 | text-ada-001 | 0.04 | #1655 |
| 5 | gpt-neox-20B | 0.05 | #1600 |
| 6 | gpt-j-6B | 0.05 | #1609 |
| 7 | OPT-175B | 0.05 | #1538 |
| 8 | OPT-66B | 0.06 | #1603 |
| 9 | GLM-130B | 0.06 | #1568 |
| 10 | text-babbage-001 | 0.06 | #1648 |
| 11 | text-davinci-002 | 0.06 | #1225 |
| 12 | text-curie-001 | 0.06 | #1632 |
| 13 | davinci | 0.07 | #1553 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=realtoxicityprompts · How It Works · Data refreshed daily, snapshot 2026-10-20.