FaithJudge: leaderboard

The model answers from given passages only (summaries, question answering, business overviews from JSON), and o3-mini (high), prompted with human hallucination annotations for the same source, flags unsupported claims; the number is the share of the model's 511 responses flagged. Unlike Vectara's HHEM leaderboard, the judge is an LLM given annotated examples rather than a fine-tuned classifier.

Metric: Hallucination rate (%), pooled over 511 RAG responses (FaithBench and RAGTruth), judged by o3-mini (high) with human-annotated examples. Source: raw.githubusercontent.com. Saturation forecast: Estimated already saturated. 35 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 Flash6.26#467
2Gemini 2.5 Pro6.65#274
3Gemini 2.5 Pro (Preview 03-25)7.63#239
4DeepSeek R1 05289.78#431
5Gemini 2.0 Flash (001)10.18#727
6GPT-4.111.94#505
7GPT-4.5 (Preview)11.94#499
8Claude Opus 4 (20250514)Thinking14.09#300
9GPT-3.5 Turbo (0125)14.87#3575
10Grok 315.26#606
11GPT-4.1 Mini15.66#793
12GPT-4o (2024-11-20)15.85#864
13Claude 3.7 Sonnet (20250219)16.05#421
14Llama 3.3 70B Instruct16.44#1387
15Phi-417.03#1925

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=faithjudge · How It Works · Data refreshed daily, snapshot 2026-11-17.