Executable Games: leaderboard
Interactive reasoning benchmark of 474 executable games (Fan et al., 2026) over sets, sequences, trees and graphs with deductive, inductive and abductive targets: the model sees only the rules, queries a hidden state turn by turn with XML-tagged actions and submits an answer. The score is the percent of games solved, averaged over five difficulty levels, with malformed actions and timeouts counted as failures; higher is better.
Metric: Success Rate (%; clean games, mean over five difficulty levels). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 74.21 | #91 |
| 2 | Gemini 2.5 Flash | 65.48 | #452 |
| 3 | GPT-5.4 | 49.17 | #121 |
| 4 | DeepSeek V3.2 | 45.49 | #403 |
| 5 | Gemini 3.1 Flash Lite | 44.08 | #363 |
| 6 | Qwen 3 Max | 43.67 | #401 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=executable-games · How It Works · Data refreshed daily, snapshot 2026-11-12.