GameReplica: leaderboard
GameReplica (Qiao et al., 2026) tests black-box game replication: a coding agent reaches one of 25 web games at five difficulty levels only through screenshots and mouse, keyboard and reset actions, and must deliver a runnable replica and a document of the rules it induced. The overall score in percent is the equal-weight mean of rule consistency, implementation consistency and visual fidelity, judged by gemini-3.6-flash with review by gpt-5.6-sol and averaged over 125 tasks and three runs; each pairing of model and agent runtime is a separate entry, higher is better.
Metric: Overall replication score (%; equal-weight mean of rule consistency, implementation consistency and visual fidelity; 125 tasks (25 web games x 5 difficulty levels), game reached only through screenshots and mouse, keyboard and reset actions; mean of 3 runs per task; first-round judge gemini-3.6-flash, review by gpt-5.6-sol). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 71.6 | #56 |
| 2 | GPT-5.5 | 42.9 | #48 |
| 3 | Grok 4.5 | 27.9 | #78 |
| 4 | GLM-5.2 | 17.8 | #137 |
| 5 | Kimi K3 | 12.7 | #53 |
| 6 | DeepSeek V4 Pro | 4 | #181 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gamereplica · How It Works · Data refreshed daily, snapshot 2026-11-12.