Swiss-Bench SBP-002 - Hallucination Detection: leaderboard
Metric: Correct rate (%) on the 63 hallucination-detection items (plausible but wrong legal claims to identify and correct) of Swiss-Bench SBP-002 (expert-written Swiss regulatory and legal questions in German, French and Italian), answered from parametric knowledge (no retrieval or system prompt, temperature 0, 4,096 output tokens); an item counts as correct when the majority grade of a three-judge panel (GPT-4o, Claude Sonnet 4, Qwen3-235B) scoring legal accuracy, citation accuracy and completeness against the reference answer is C (weighted score at least 0.8); higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 Plus (2026-02-15) | 30.2 | #114 |
| 2 | Gemini 2.5 Flash | 15.9 | #237 |
| 3 | GLM-5 | 9.5 | #137 |
| 4 | GPT-4o | 6.3 | #333 |
| 5 | Claude Sonnet 4 | 6.3 | #194 |
| 6 | Mistral Large 2 (Nov) Instruct (2411) | 6.3 | #458 |
| 7 | MiniMax-M2.5 | 4.8 | #295 |
| 8 | GPT-OSS-120B | 1.6 | #330 |
| 9 | MiMo-V2-Flash | 0 | #359 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=swiss-bench-sbp-002-hallucination-detection · How It Works · Data refreshed daily, snapshot 2026-10-11.