CubeBench: leaderboard
Agentic spatial reasoning benchmark using Rubik's Cube solving. Models interact with a simulated cube through tool calls over up to 50 turns, tested on 5 task families from cross to full solve.
Metric: Success Rate (%). Source: github.com. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 66.7 | #79 (GPT-5.4) |
| 2 | Gemini 3.1 Pro (Preview) (High) | 55.6 | #58 (Gemini 3.1 Pro (Preview)) |
Interactive version: theaggregate.ai/benchmark?slug=cubebench · How It Works · Data refreshed daily, snapshot 2026-10-20.