Skip to main content
Meo

Theory of mind

Reasoning about other agents’ beliefs and knowledge. As of 2026-10-18, the strongest model on this domain is OpenAI: GPT-5.5 (100% — 3 of 3 questions correct).

What this domain tests

Tracking what other agents know, believe, and can infer — including false-belief tasks where an agent’s view diverges from reality. Tests social/recursive reasoning, not factual recall.

Why it resists contamination

Newly authored scenarios on a private holdout, graded by the multi-lab jury, so canonical false-belief stories cannot be answered from memory. See the full methodology.

Best AI models for Theory of mind

#ModelTheory of mind accuracyQuestionsOverall𝕍 index
1OpenAI: GPT-5.5100%3/394.1%23.2
2Google: Gemini 4 Argon100%3/392.5%18.2
3Anthropic: Claude Opus 5.5100%3/392.1%100.0
4OpenAI: GPT-5.5 Pro100%3/391.6%4.3
5OpenAI: GPT-6.1 Sol100%3/390.4%35.6
6Anthropic: Claude Fable 5100%3/388.7%27.5
7Anthropic: Claude Opus 4.8100%3/384.9%12.1
8inclusionAI: Ring-2.6-1T100%3/384.7%14.8
9Z.ai: GLM 5.2100%3/383.2%0.78
10Google: Gemini 3.5 Flash100%3/382.0%11.1
11DeepSeek: DeepSeek V4 Flash 0423100%3/381.8%0.43
12Qwen: Qwen3.7 Max100%3/381.0%0.65
13MoonshotAI: Kimi K3100%3/380.3%3.6
14DeepSeek: DeepSeek V4 Pro 0423100%3/380.3%0.53
15Google: Gemini 3.1 Pro Preview100%3/379.5%3.2
16NVIDIA: Nemotron 3 Ultra100%3/375.2%1.2
17Owl Alpha100%3/373.7%0.37
18SpaceXAI: Grok 4.7100%3/372.4%0.17
19Xiaomi: MiMo-V2.5-Pro100%3/371.5%0.18
20OpenAI: GPT-5.4 Nano100%3/369.5%2.3
21Qwen: Qwen3.6 Flash100%3/369.5%0.17
22Qwen: Qwen3.5-Flash100%3/368.2%0.15
23MiniMax: MiniMax M3100%3/367.4%0.13
24MoonshotAI: Kimi K2.6100%3/366.9%0.050
25MoonshotAI: Kimi K2.7 Code100%3/366.9%0.073
26OpenAI: GPT-5 Mini100%3/366.9%0.25
27Qwen: Qwen3.7 Plus100%3/366.9%0.051
28OpenAI: GPT-5.1-Codex-Mini100%3/366.5%1.3
29SpaceXAI: Grok 4.3100%3/366.5%1.7
30Xiaomi: MiMo-V2.5100%3/366.5%0.26
31Qwen: Qwen3 VL 32B Instruct100%3/366.1%0.19
32Nex AGI: Nex-N2-Pro (free)100%3/365.7%0.18
33ByteDance Seed: Seed 1.6100%3/364.9%0.036
34Qwen: Qwen3 VL 235B A22B Instruct100%3/364.4%0.060
35Qwen: Qwen3 VL 30B A3B Thinking100%3/364.0%0.008
36Google: Gemma 4 31B100%3/363.6%0.009
37Z.ai: GLM 5.1100%3/363.5%0.033
38Baidu: ERNIE 4.5 VL 424B A47B 100%3/363.2%0.015
39Google: Gemini 3.1 Flash Lite Preview100%3/361.9%0.46
40Qwen: Qwen3.5 Plus 2026-02-15100%3/361.9%0.012
41Mistral: Mistral Large 3 2512100%3/359.0%0.17
42Google: Gemini 2.5 Flash Lite100%3/357.3%0.13
43Google: Gemini 3.1 Flash Lite100%3/356.5%0.13
44Google: Gemma 4 26B A4B 100%3/355.2%0.003
45Meta: Llama 4 Maverick100%3/352.7%0.008
46Qwen: Qwen2.5 VL 72B Instruct100%3/350.2%0.005
47StepFun: Step 3.7 Flash100%3/349.8%0.005
48Qwen: Qwen3.5-27B100%3/349.4%<0.001
49Qwen: Qwen3.6 35B A3B100%3/347.7%0.004
50Qwen: Qwen3.6 Plus100%3/344.8%<0.001
51Qwen: Qwen3 VL 30B A3B Instruct67%2/366.1%0.13
52Z.ai: GLM 4.6V67%2/365.7%0.12
53OpenAI: GPT-4.1 Mini67%2/363.6%0.26
54ByteDance Seed: Seed 1.6 Flash67%2/357.7%0.080
55Perceptron: Perceptron Mk167%2/357.7%0.055
56Mistral: Ministral 3 14B 251267%2/356.5%0.086
57Tencent: Hy3 preview67%2/354.0%0.007
58OpenAI: GPT-5 Nano67%2/354.0%0.026
59Mistral: Mistral Small 3.2 24B67%2/353.6%0.017
60Google: Gemma 3 27B67%2/350.2%0.006
61MiniMax: MiniMax-0167%2/349.4%0.005
62Google: Gemma 3 12B67%2/349.0%0.009
63NVIDIA: Nemotron 3 Nano Omni (free)67%2/348.5%0.016
64ByteDance Seed: Seed-2.0-Lite67%2/347.7%<0.001
65OpenAI: GPT-4o-mini67%2/347.7%0.021
66Qwen: Qwen3.5-9B67%2/341.4%<0.001
67Qwen: Qwen3.5-35B-A3B67%2/339.3%<0.001
68ByteDance Seed: Seed-2.0-Mini67%2/337.2%<0.001
69Meta: Llama 3.2 11B Vision Instruct67%2/335.6%<0.001
70Qwen: Qwen3 VL 8B Thinking33%1/363.2%0.072
71Mistral: Mistral Small 433%1/357.7%0.065
72Qwen: Qwen3 VL 8B Instruct33%1/356.9%0.028
73Meta: Llama 4 Scout33%1/348.5%0.009
74Arcee AI: Trinity Large Thinking33%1/341.6%<0.001
75Amazon: Nova Lite 1.033%1/339.7%0.012
76Mistral: Ministral 3 8B 25120%0/354.0%0.065
77OpenAI: GPT-4.1 Nano0%0/348.1%0.044
78Google: Gemma 3 4B0%0/347.3%0.003
79Mistral: Ministral 3 3B 25120%0/345.6%0.021
80Perplexity: Sonar0%0/345.6%0.004
81NVIDIA: Nemotron Nano 12B 2 VL (free)0%0/342.3%0.002
82Anthropic: Claude 3 Haiku0%0/341.8%0.014

Primary sources & further reading

Frequently asked questions

What does the meo Theory of mind benchmark test?

Tracking what other agents know, believe, and can infer — including false-belief tasks where an agent’s view diverges from reality. Tests social/recursive reasoning, not factual recall.

Why can't AI models game the Theory of mind benchmark?

Newly authored scenarios on a private holdout, graded by the multi-lab jury, so canonical false-belief stories cannot be answered from memory.

Which AI model is best at Theory of mind?

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

Related: all domains · Theory of mind ranking · methodology · full leaderboard.