Benchmarks · safety
Spiral-Bench
A LLM-judged benchmark measuring sycophancy and delusion reinforcement in simulated multi-turn chats. Maintained by sam-paech. Score: Safety Score: weighted average of protective vs risky behaviours, scaled 0-100.
21 models · 21 results · data as of 2026-10-29 · eqbench.com ↗
Leaderboard
GPT-5 high
All results
21 rows| # | Model | Effort | Spiral-Bench |
|---|---|---|---|
| 1 | GPT-5 | high | 70.8% |
| 2 | Sonnet 4.5 | high | 70.3% |
| 3 | GPT-5.2 | xhigh | 70.2% |
| 4 | o3 | high | 63.4% |
| 5 | gpt-oss-120b | high | 57.5% |
| 6 | GPT-5.2025 | high | 57.4% |
| 7 | Kimi K2 | medium | 57.2% |
| 8 | o4 mini | high | 53.6% |
| 9 | Sonnet 3.5 | medium | 42.6% |
| 10 | Gemini 3 Pro | high | 38.7% |
| 11 | Llama 4 Maverick | medium | 38.2% |
| 12 | Grok 4 Fast | none | 37.5% |
| 13 | Gemini 2.5 Flash | medium | 36.5% |
| 14 | Sonnet 4 | high | 32.1% |
| 15 | GLM 4.6 | none | 30.8% |
| 16 | Gemini 2.5 Pro | high | 28.3% |
| 17 | Grok 4.1 Fast | high | 25.3% |
| 18 | Gemma 3 27B IT | medium | 25.2% |
| 19 | Qwen 3 235B-A22B | xhigh | 20.4% |
| 20 | Mistral Medium 3.1 | none | 15.0% |
| 21 | DeepSeek R1.0528 | 14.2% |