Skip to main content
Meo

Tape-machine simulation

Executing a small abstract machine step by step. As of 2026-10-18, the strongest model on this domain is OpenAI: GPT-5.5 (100% — 15 of 15 questions correct).

What this domain tests

Simulating a small tape/register machine for N steps and reporting the resulting state — faithful, long-horizon execution without losing the thread.

Why it resists contamination

The machine, its program, and the correct end-state are generated and simulated programmatically, so the answer key exists by construction (generator-as-oracle). Because instances are generated with a programmatic ground truth, this is a “generator-as-oracle” domain — effectively unlimited fresh, self-verifying items. See the full methodology.

Best AI models for Tape-machine simulation

#ModelTape-machine simulation accuracyQuestionsOverall𝕍 index
1OpenAI: GPT-5.5100%15/1594.1%23.2
2Google: Gemini 4 Argon100%15/1592.5%18.2
3OpenAI: GPT-5.5 Pro100%15/1591.6%4.3
4OpenAI: GPT-6.1 Sol100%15/1590.4%35.6
5Anthropic: Claude Opus 4.8100%15/1584.9%12.1
6Google: Gemini 3.5 Flash100%15/1582.0%11.1
7DeepSeek: DeepSeek V4 Flash 0423100%15/1581.8%0.43
8Qwen: Qwen3.7 Max100%15/1581.0%0.65
9MoonshotAI: Kimi K3100%15/1580.3%3.6
10Google: Gemini 3.1 Pro Preview100%15/1579.5%3.2
11NVIDIA: Nemotron 3 Ultra100%15/1575.2%1.2
12SpaceXAI: Grok 4.7100%15/1572.4%0.17
13Xiaomi: MiMo-V2.5-Pro100%15/1571.5%0.18
14OpenAI: GPT-5.4 Nano100%15/1569.5%2.3
15Qwen: Qwen3.6 Flash100%15/1569.5%0.17
16Qwen: Qwen3.5-Flash100%15/1568.2%0.15
17MoonshotAI: Kimi K2.6100%15/1566.9%0.050
18OpenAI: GPT-5 Mini100%15/1566.9%0.25
19OpenAI: GPT-5.1-Codex-Mini100%15/1566.5%1.3
20SpaceXAI: Grok 4.3100%15/1566.5%1.7
21Qwen: Qwen3 VL 32B Instruct100%15/1566.1%0.19
22ByteDance Seed: Seed 1.6100%15/1564.9%0.036
23Qwen: Qwen3 VL 235B A22B Instruct100%15/1564.4%0.060
24Google: Gemma 4 31B100%15/1563.6%0.009
25OpenAI: GPT-4.1 Mini100%15/1563.6%0.26
26Qwen: Qwen3.5 Plus 2026-02-15100%15/1561.9%0.012
27Mistral: Mistral Large 3 2512100%15/1559.0%0.17
28Perceptron: Perceptron Mk1100%15/1557.7%0.055
29Anthropic: Claude Opus 5.593%14/1592.1%100.0
30Anthropic: Claude Fable 593%14/1588.7%27.5
31inclusionAI: Ring-2.6-1T93%14/1584.7%14.8
32Z.ai: GLM 5.293%14/1583.2%0.78
33DeepSeek: DeepSeek V4 Pro 042393%14/1580.3%0.53
34MiniMax: MiniMax M393%14/1567.4%0.13
35MoonshotAI: Kimi K2.7 Code93%14/1566.9%0.073
36Qwen: Qwen3.7 Plus93%14/1566.9%0.051
37Nex AGI: Nex-N2-Pro (free)93%14/1565.7%0.18
38Qwen: Qwen3 VL 30B A3B Thinking93%14/1564.0%0.008
39Google: Gemini 3.1 Flash Lite Preview93%14/1561.9%0.46
40Mistral: Ministral 3 8B 251293%14/1554.0%0.065
41Mistral: Mistral Small 3.2 24B93%14/1553.6%0.017
42Qwen: Qwen3.5-27B93%14/1549.4%<0.001
43Owl Alpha87%13/1573.7%0.37
44Xiaomi: MiMo-V2.587%13/1566.5%0.26
45Qwen: Qwen3 VL 8B Thinking87%13/1563.2%0.072
46Google: Gemini 2.5 Flash Lite87%13/1557.3%0.13
47Mistral: Ministral 3 14B 251287%13/1556.5%0.086
48Google: Gemma 3 27B87%13/1550.2%0.006
49Meta: Llama 4 Scout87%13/1548.5%0.009
50Baidu: ERNIE 4.5 VL 424B A47B 80%12/1563.2%0.015
51Google: Gemma 4 26B A4B 80%12/1555.2%0.003
52Meta: Llama 4 Maverick80%12/1552.7%0.008
53MiniMax: MiniMax-0180%12/1549.4%0.005
54Qwen: Qwen3 VL 30B A3B Instruct73%11/1566.1%0.13
55Qwen: Qwen2.5 VL 72B Instruct73%11/1550.2%0.005
56Z.ai: GLM 4.6V67%10/1565.7%0.12
57ByteDance Seed: Seed 1.6 Flash67%10/1557.7%0.080
58Qwen: Qwen3 VL 8B Instruct67%10/1556.9%0.028
59Google: Gemma 3 12B67%10/1549.0%0.009
60OpenAI: GPT-4o-mini67%10/1547.7%0.021
61Anthropic: Claude 3 Haiku67%10/1541.8%0.014
62Z.ai: GLM 5.153%8/1563.5%0.033
63Mistral: Mistral Small 453%8/1557.7%0.065
64OpenAI: GPT-5 Nano53%8/1554.0%0.026
65OpenAI: GPT-4.1 Nano53%8/1548.1%0.044
66Mistral: Ministral 3 3B 251253%8/1545.6%0.021
67Google: Gemma 3 4B40%6/1547.3%0.003
68Meta: Llama 3.2 11B Vision Instruct40%6/1535.6%<0.001
69Tencent: Hy3 preview33%5/1554.0%0.007
70Perplexity: Sonar33%5/1545.6%0.004
71Amazon: Nova Lite 1.033%5/1539.7%0.012
72Google: Gemini 3.1 Flash Lite27%4/1556.5%0.13
73ByteDance Seed: Seed-2.0-Lite27%4/1547.7%<0.001
74NVIDIA: Nemotron 3 Nano Omni (free)7%1/1548.5%0.016
75Qwen: Qwen3.6 Plus7%1/1544.8%<0.001
76Qwen: Qwen3.5-35B-A3B7%1/1539.3%<0.001
77StepFun: Step 3.7 Flash0%0/1549.8%0.005
78Qwen: Qwen3.6 35B A3B0%0/1547.7%0.004
79NVIDIA: Nemotron Nano 12B 2 VL (free)0%0/1542.3%0.002
80Arcee AI: Trinity Large Thinking0%0/1541.6%<0.001
81Qwen: Qwen3.5-9B0%0/1541.4%<0.001
82ByteDance Seed: Seed-2.0-Mini0%0/1537.2%<0.001

Frequently asked questions

What does the meo Tape-machine simulation benchmark test?

Simulating a small tape/register machine for N steps and reporting the resulting state — faithful, long-horizon execution without losing the thread.

Why can't AI models game the Tape-machine simulation benchmark?

The machine, its program, and the correct end-state are generated and simulated programmatically, so the answer key exists by construction (generator-as-oracle).

Which AI model is best at Tape-machine simulation?

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

Related: all domains · Tape-machine simulation ranking · methodology · full leaderboard.