Benchmark
Spiral-Bench
Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages.
All safety benchmarksScores from EQ-Bench boards.
- Models measured
- 22
- Scores recorded
- 22
- Sources
- 1
- IQ weight
- 0.64
Score distribution
- 41.4%
- Mean
- 38.2%
- Median
- 17.9%
- Std dev
- 20.4%
- 10th percentile
- 70.2%
- 90th percentile
- 14.2%
- Min
This benchmark has a weight of 0.64 for chat models, measured over 22 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 Chat· EQ-Bench boards | 70.8% |
| 2 | Claude Sonnet 4.5· EQ-Bench boards | 70.3% |
| 3 | GPT 5.2· EQ-Bench boards | 70.2% |
| 4 | OpenAI o3· EQ-Bench boards | 63.4% |
| 5 | GPT OSS 120B· EQ-Bench boards | 57.5% |
| 6 | GPT 5· EQ-Bench boards | 57.4% |
| 7 | Kimi K2· EQ-Bench boards | 57.2% |
| 8 | o4 Mini· EQ-Bench boards | 53.6% |
| 9 | Claude 3.5 Sonnet· EQ-Bench boards | 42.6% |
| 10 | Gemini 3 Pro Preview· EQ-Bench boards | 38.7% |
| 11 | Llama 4 Maverick· EQ-Bench boards | 38.2% |
| 12 | Grok 4 Fast· EQ-Bench boards | 37.5% |
| 13 | Gemini 2.5 Flash· EQ-Bench boards | 36.5% |
| 14 | Claude Sonnet 4· EQ-Bench boards | 32.1% |
| 15 | GLM 4.6· EQ-Bench boards | 30.8% |
| 16 | Gemini 2.5 Pro· EQ-Bench boards | 28.3% |
| 17 | Grok 4.1 Fast· EQ-Bench boards | 25.3% |
| 18 | ChatGPT-4o Latest· EQ-Bench boards | 25.2% |
| 19 | Gemma 3 27B IT· EQ-Bench boards | 25.2% |
| 20 | Qwen3 235B A22B Instruct 2507· EQ-Bench boards | 20.4% |
| 21 | Mistral Medium 3.1· EQ-Bench boards | 15% |
| 22 | DeepSeek R1· EQ-Bench boards | 14.2% |
Data credits: EQ-Bench (Samuel J. Paech) — cite paech2023eqbench