SWE-PRBench - Hallucination Rate: leaderboard
Metric: Hallucination rate (%): share of the model's review comments the judge classes as fabricated (code not in context or factually wrong claims), on a stratified 100-PR sample of SWE-PRBench (40 direct, 40 contextual and 20 latent-difficulty pull requests from active open-source repositories), each model asked at temperature 0 for 4 to 6 line-anchored review comments as JSON, a GPT-5.2 judge matching them to the human reviewers' comments (confirmed, plausible or fabricated), averaged over the three context configurations; lower is better. Source: arxiv.org. Saturation forecast: Around January 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 19.3 | #333 |
| 2 | Claude Sonnet 4.6 | 22.7 | #85 |
| 3 | Mistral Small | 25.1 | #700 |
| 4 | DeepSeek V3 | 31.5 | #312 |
| 5 | Claude Haiku 4.5 | 34.6 | #271 |
| 6 | GPT-4o Mini | 35.3 | #588 |
| 7 | Mistral Large 3 | 35.3 | #388 |
| 8 | Llama 3.3 70B | 41.7 | #569 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=swe-prbench-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-10-11.