Skip to main content
Meo

Perceptual illusions

Questions engineered to trigger a confident wrong answer. As of 2026-10-18, the strongest model on this domain is OpenAI: GPT-5.5 (100% — 102 of 102 questions correct).

What this domain tests

Cognitive/perceptual “illusion” problems where the obvious first answer is wrong — measuring whether a model overrides a strong but incorrect prior on reflection.

Why it resists contamination

Freshly authored variants on a private holdout, graded by the multi-lab jury, so the trick cannot be answered from memory of the canonical version. See the full methodology.

Best AI models for Perceptual illusions

#ModelPerceptual illusions accuracyQuestionsOverall𝕍 index
1OpenAI: GPT-5.5100%102/10294.1%23.2
2Anthropic: Claude Opus 5.5100%102/10292.1%100.0
3OpenAI: GPT-6.1 Sol100%102/10290.4%35.6
4Google: Gemini 4 Argon99%101/10292.5%18.2
5OpenAI: GPT-5.5 Pro98%100/10291.6%4.3
6Anthropic: Claude Fable 589%91/10288.7%27.5
7Google: Gemini 3.5 Flash80%82/10282.0%11.1
8Anthropic: Claude Opus 4.878%80/10284.9%12.1
9Google: Gemini 3.1 Pro Preview75%76/10279.5%3.2
10MoonshotAI: Kimi K370%71/10280.3%3.6
11Z.ai: GLM 4.6V63%64/10265.7%0.12
12Meta: Llama 4 Maverick63%64/10252.7%0.008
13SpaceXAI: Grok 4.759%60/10272.4%0.17
14OpenAI: GPT-4.1 Nano59%60/10248.1%0.044
15Baidu: ERNIE 4.5 VL 424B A47B 58%59/10263.2%0.015
16Google: Gemma 3 12B58%59/10249.0%0.009
17Google: Gemma 3 4B57%58/10247.3%0.003
18Perceptron: Perceptron Mk156%57/10257.7%0.055
19Google: Gemini 2.5 Flash Lite56%57/10257.3%0.13
20StepFun: Step 3.7 Flash56%57/10249.8%0.005
21OpenAI: GPT-4o-mini56%57/10247.7%0.021
22NVIDIA: Nemotron Nano 12B 2 VL (free)56%57/10242.3%0.002
23Qwen: Qwen3.5-9B56%57/10241.4%<0.001
24Qwen: Qwen3.7 Plus55%56/10266.9%0.051
25OpenAI: GPT-4.1 Mini55%56/10263.6%0.26
26Meta: Llama 4 Scout55%56/10248.5%0.009
27MiniMax: MiniMax M354%55/10267.4%0.13
28SpaceXAI: Grok 4.354%55/10266.5%1.7
29Qwen: Qwen3 VL 30B A3B Instruct54%55/10266.1%0.13
30Google: Gemini 3.1 Flash Lite Preview54%55/10261.9%0.46
31MiniMax: MiniMax-0154%55/10249.4%0.005
32ByteDance Seed: Seed-2.0-Lite54%55/10247.7%<0.001
33Qwen: Qwen3.6 35B A3B54%55/10247.7%0.004
34Qwen: Qwen3.6 Flash53%54/10269.5%0.17
35Qwen: Qwen3 VL 32B Instruct53%54/10266.1%0.19
36ByteDance Seed: Seed 1.653%54/10264.9%0.036
37Google: Gemini 3.1 Flash Lite53%54/10256.5%0.13
38Anthropic: Claude 3 Haiku53%54/10241.8%0.014
39Qwen: Qwen3.5-Flash52%53/10268.2%0.15
40MoonshotAI: Kimi K2.7 Code52%53/10266.9%0.073
41OpenAI: GPT-5 Mini52%53/10266.9%0.25
42OpenAI: GPT-5.1-Codex-Mini52%53/10266.5%1.3
43Xiaomi: MiMo-V2.552%53/10266.5%0.26
44Qwen: Qwen3 VL 30B A3B Thinking52%53/10264.0%0.008
45Qwen: Qwen3.6 Plus52%53/10244.8%<0.001
46MoonshotAI: Kimi K2.651%52/10266.9%0.050
47Nex AGI: Nex-N2-Pro (free)51%52/10265.7%0.18
48Qwen: Qwen3 VL 235B A22B Instruct51%52/10264.4%0.060
49Qwen: Qwen3.5 Plus 2026-02-1551%52/10261.9%0.012
50Mistral: Mistral Large 3 251251%52/10259.0%0.17
51Mistral: Mistral Small 451%52/10257.7%0.065
52Qwen: Qwen2.5 VL 72B Instruct51%52/10250.2%0.005
53Qwen: Qwen3.5-27B51%52/10249.4%<0.001
54ByteDance Seed: Seed-2.0-Mini51%52/10237.2%<0.001
55ByteDance Seed: Seed 1.6 Flash50%51/10257.7%0.080
56Qwen: Qwen3 VL 8B Instruct50%51/10256.9%0.028
57Mistral: Ministral 3 8B 251250%51/10254.0%0.065
58Perplexity: Sonar50%51/10245.6%0.004
59Mistral: Mistral Small 3.2 24B49%50/10253.6%0.017
60NVIDIA: Nemotron 3 Nano Omni (free)49%50/10248.5%0.016
61OpenAI: GPT-5.4 Nano48%49/10269.5%2.3
62OpenAI: GPT-5 Nano48%49/10254.0%0.026
63Google: Gemma 3 27B48%49/10250.2%0.006
64Mistral: Ministral 3 3B 251248%49/10245.6%0.021
65Amazon: Nova Lite 1.048%49/10239.7%0.012
66Qwen: Qwen3.5-35B-A3B48%49/10239.3%<0.001
67Google: Gemma 4 26B A4B 47%48/10255.2%0.003
68Meta: Llama 3.2 11B Vision Instruct47%48/10235.6%<0.001
69Google: Gemma 4 31B46%47/10263.6%0.009
70Qwen: Qwen3 VL 8B Thinking46%47/10263.2%0.072
71Mistral: Ministral 3 14B 251245%46/10256.5%0.086

Frequently asked questions

What does the meo Perceptual illusions benchmark test?

Cognitive/perceptual “illusion” problems where the obvious first answer is wrong — measuring whether a model overrides a strong but incorrect prior on reflection.

Why can't AI models game the Perceptual illusions benchmark?

Freshly authored variants on a private holdout, graded by the multi-lab jury, so the trick cannot be answered from memory of the canonical version.

Which AI model is best at Perceptual illusions?

OpenAI: GPT-5.5 leads the Perceptual illusions domain on the meo benchmark with 100% accuracy (as of 2026-10-18).

Related: all domains · Perceptual illusions ranking · methodology · full leaderboard.