Skip to content

Benchmark

PerceptionBench · Perception-related hallucination accuracy

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

Models and systems measured
15
Scores recorded
15
Sources
1
IQ weight
0.39

Score distribution

35.8%
Mean
34.7%
Median
7.6%
Std dev
26.9%
10th percentile
45%
90th percentile
26.9%
Min
26.9%50.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.39 for chat models, measured over 11 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model or systemScore
1Gemini 3.5 Flashbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · high reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs50.6%
2Claude Fable 5benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · max reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs45%
3Grok 4.5benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs44.7%
4Kimi K3benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · max reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.41.7%
5Kimi K2.6benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.41%
6Gemini 3.1 Pro Previewbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · high reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs40.6%
7Claude Opus 4.8benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · max reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs38.7%
8GPT 5.5benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · xhigh reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs34.7%
9MiniMax M3benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs29.9%
10Qwen3.7 Plusbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs29.5%
11GPT-5.6 Solbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · max reasoning · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs26.9%
12Qwen3.5 397B A17Bbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs26.9%

Not ranked

The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.

Model or systemScore
Gemma 4 31B (PerceptionBench label)benchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · thinking mode enabled · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs32.1%
GLM 5V Turbobenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs28%
GLM-4.6Vbenchmark-owner evaluation · 3,000 verified questions total · open-ended short answers · unified prompts · highest available reasoning budget; level not specified · main-table result; run count not specified · GPT-oss-120B judge· Reported by Moonshot AI / PerceptionBench authors· via BenchBoard docs26.9%

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.