Skip to content

Benchmark

HalluHard Hallucination Rate (research questions)

Factuality benchmarks measure how often models get facts right. Scores are percentages. Lower is better.

Models measured
30
Scores recorded
40
Sources
1
IQ weight
2.56

Score distribution

79%
Mean
84.1%
Median
16.4%
Std dev
54.8%
10th percentile
92.9%
90th percentile
29.6%
Min
29.6%97.4%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.56 for chat models, measured over 30 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude Opus 4.5websearch· HalluHard Leaderboard data docs29.6%
2GLM 5websearch, thinking· HalluHard Leaderboard data docs38.9%
3GPT-6 Astrawebsearch· HalluHard Leaderboard data docs44.4%
4GPT-5.3websearch· HalluHard Leaderboard data docs54%
5GPT 5.4websearch, thinking· HalluHard Leaderboard data docs54.8%
6Claude Fable 5websearch· HalluHard Leaderboard data docs54.9%
7Kimi K2.5websearch· HalluHard Leaderboard data docs63.4%
8GPT-6 Astra· HalluHard Leaderboard data docs66.7%
9GPT-5.3· HalluHard Leaderboard data docs69.1%
10Claude Fable 5· HalluHard Leaderboard data docs71.1%
11GPT 5.4thinking· HalluHard Leaderboard data docs72.7%
12GPT 5.2thinking· HalluHard Leaderboard data docs75.3%
13Gemini 3.1 Pro Previewwebsearch· HalluHard Leaderboard data docs77.2%
14Gemini 3.1 Pro Preview· HalluHard Leaderboard data docs78.2%
15GPT 5.2· HalluHard Leaderboard data docs79.4%
16Claude Sonnet 5· HalluHard Leaderboard data docs80.1%
17Claude Opus 4.6· HalluHard Leaderboard data docs81.6%
18Kimi K2.6thinking· HalluHard Leaderboard data docs82.9%
19Claude Opus 4.7· HalluHard Leaderboard data docs83.8%
20Claude Opus 4.5· HalluHard Leaderboard data docs84%
21Gemini 3 Pro Preview· HalluHard Leaderboard data docs84.1%
22Claude Sonnet 4.6· HalluHard Leaderboard data docs85.8%
23Claude Sonnet 4.5· HalluHard Leaderboard data docs87.3%
24GPT 5thinking· HalluHard Leaderboard data docs87.3%
25DeepSeek Reasoner· HalluHard Leaderboard data docs88.6%
26Gemini 3 Flash· HalluHard Leaderboard data docs88.6%
27GLM 5.2· HalluHard Leaderboard data docs89.7%
28DeepSeek Chat· HalluHard Leaderboard data docs90.1%
29DeepSeek V4 Pro· HalluHard Leaderboard data docs90.1%
30GLM 5thinking· HalluHard Leaderboard data docs90.1%
31GLM 4.7thinking· HalluHard Leaderboard data docs90.7%
32GPT 5· HalluHard Leaderboard data docs91.1%
33GPT-5 mini· HalluHard Leaderboard data docs92.6%
34Grok 4thinking· HalluHard Leaderboard data docs92.8%
35Claude Haiku 4.5· HalluHard Leaderboard data docs92.9%
36Kimi K2.5thinking· HalluHard Leaderboard data docs92.9%
37Kimi K2 Thinking· HalluHard Leaderboard data docs93.5%
38Nemotron 3 Ultra· HalluHard Leaderboard data docs93.7%
39GPT-5 nano· HalluHard Leaderboard data docs96.9%
40Grok 4.1 Fastthinking· HalluHard Leaderboard data docs97.4%

Data credits: HalluHard: Fan, Delsad, Flammarion and Andriushchenko (EPFL), arXiv:2602.01031