ERR-EVAL: leaderboard
ERR-EVAL scores epistemic reliability: every prompt is defective on purpose — garbled, ambiguous, false-premised, missing a fact, or impossible — and a model is graded on whether it notices, says which part it cannot answer, and declines to invent the rest. Higher is better.
Metric: Overall score. Source: err-eval.gustycube.com. Saturation forecast: Around April 2027. 60 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 6.4 | #115 |
| 2 | Claude Sonnet 4.6 | 5.88 | #94 |
| 3 | Claude Opus 4.5 | 5.8 | #87 |
| 4 | GPT-5.2 Pro | 5.76 | #70 |
| 5 | GPT-5.4 | 5.6 | #83 |
| 6 | Claude Opus 4.6 | 5.48 | #67 |
| 7 | Claude Sonnet 4.5 | 5.28 | #148 |
| 8 | Claude Haiku 4.5 | 5.04 | #287 |
| 9 | GPT-5.3 Codex | 5 | #72 |
| 10 | GPT-5.4 Mini | 4.96 | #228 |
| 11 | GPT-5.1 | 4.84 | #145 |
| 12 | GPT-5.3 | 4.76 | #104 |
| 13 | GPT-5 Mini | 4.68 | #193 |
| 14 | Grok 4 | 4.68 | #189 |
| 15 | Claude Sonnet 4 | 4.64 | #207 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=err-eval · How It Works · Data refreshed daily, snapshot 2026-11-05.