CiteTrace - Fidelity Failure Rate: leaderboard
How often search-augmented LLM APIs misrepresent the pages they cite, on CiteTrace (Seo et al., 2026, Verified Misguidance): 11,200 real Stack Exchange questions from 28 communities, every cited URL crawled and each citing sentence checked against its source by a gpt-4o-mini judge validated against human annotators. Percent of evaluable citations whose claim is fabricated or misattributed (fidelity levels 1-2 of five); lower is better. Each API answers with its provider web-search tool, so the rate describes the search-augmented product rather than the bare model.
Metric: Fidelity failure rate (%; share of evaluable citations whose cited claim is fabricated or misattributed relative to the crawled source, answer-source fidelity 1-2 on a five-level scale; gpt-4o-mini judge; 11,200 Stack Exchange questions). Source: arxiv.org. Saturation forecast: Around May 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 12.31 | #568 |
| 2 | Claude Sonnet 4.6 | 13.07 | #158 |
| 3 | Gemini 3 Flash | 29.97 | #159 |
| 4 | Gemini 3.1 Pro (Preview) | 33.07 | #91 |
| 5 | Sonar | 34.07 | #1677 |
| 6 | Grok 4.1 Fast (Non-reasoning) | 40.5 | #426 (Grok 4.1 Fast) |
| 7 | GPT-5 | 42.29 | #163 |
| 8 | Grok 4.1 Fast (Reasoning) | 42.64 | #426 (Grok 4.1 Fast) |
| 9 | GPT-5 Mini | 44.85 | #349 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=citetrace-fidelity-failure-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.