HalluHard: A Hard Multi-Turn Hallucination Benchmark: leaderboard
HalluHard evaluates text-only multi-turn hallucination behavior in large language models across legal, research, medical, and coding prompts. Models are judged on groundedness, with inline citations required for factual claims, and the pinned metric is average hallucination rate, reported as a percentage; lower is better. Source: BenchmarkList’s attributed result table.
Metric: Avg (↓) (%). Source: benchmarklist.com. Saturation forecast: Around February 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 30.2 | #87 |
| 2 | GPT-5.2 (Thinking) | 38.2 | #115 (GPT-5.2) |
| 3 | GPT-5.2 | 58.8 | #115 |
| 4 | Gemini 3 Pro | 61.9 | #82 |
| 5 | GPT-5 (Thinking) | 64.8 | #105 (GPT-5) |
| 6 | Claude Sonnet 4.5 | 65.6 | #148 |
| 7 | Gemini 3 Flash | 69.5 | #103 |
| 8 | GPT-5 | 71.8 | #105 |
| 9 | DeepSeek V3 Chat | 75.8 | #457 |
| 10 | GPT-5 Mini | 75.9 | #193 |
| 11 | DeepSeek Reasoner | 76.8 | #254 |
| 12 | GLM-4.7 (Thinking) | 77.1 | #202 (GLM-4.7) |
| 13 | Claude Haiku 4.5 | 79.5 | #287 |
| 14 | Kimi K2 (Thinking) | 80.1 | #252 (Kimi K2) |
| 15 | GPT-5 Nano | 85.1 | #439 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=halluhard-a-hard-multi-turn-hallucination-benchmark · How It Works · Data refreshed daily, snapshot 2026-11-05.