Confabulation Leaderboard (Lechmazur): leaderboard
Lechmazur's confabulation leaderboard — measures made-up answers (confabulations/hallucinations) on provably unanswerable questions versus correct refusals.
Metric: Confabulation rate % (lower is better). Source: lechmazur.github.io. Saturation forecast: Estimated already saturated. 66 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4 (Thinking 16K) | 2.48 | #163 (Claude Opus 4) |
| 2 | Claude Opus 4 (Non-reasoning) | 2.97 | #163 (Claude Opus 4) |
| 3 | Gemini 2.5 Pro | 3.96 | #152 |
| 4 | Grok 4 | 3.96 | #181 |
| 5 | Gemini 2.5 Pro (Preview 03-25) | 3.96 | #123 |
| 6 | Claude Sonnet 4 (Non-reasoning) | 5.45 | #199 (Claude Sonnet 4) |
| 7 | Gemini 2.5 Pro (Preview 05-06) | 5.94 | #147 |
| 8 | Grok 3 Mini (High) | 6.93 | #281 (Grok 3 Mini) |
| 9 | GLM-4.5 | 7.92 | #275 |
| 10 | O1 | 10.89 | #183 |
| 11 | Grok 3 Mini (Low) | 10.89 | #281 (Grok 3 Mini) |
| 12 | GPT-5 (Medium) | 10.9 | #98 (GPT-5) |
| 13 | GPT-4.5 (Preview) | 11.88 | #244 |
| 14 | Claude 3.5 Sonnet (20241022) | 12.87 | #291 |
| 15 | DeepSeek R1 0528 | 12.87 | #222 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=confabulation-leaderboard-lechmazur · How It Works · Data refreshed daily, snapshot 2026-10-20.