Step Game (Lechmazur) — leaderboard
Tests spatial reasoning - models must track relative positions of characters over 10 steps (e.g., 'A is to the left of B'). TrueSkill ranking from multi-game tournaments across 75+ models.
Metric: TrueSkill μ. Source: github.com. Status: saturation imminent. 75 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 (Medium) | 5.49 |
| 2 | O3 (Medium) | 5.32 |
| 3 | GPT-5.1 (Medium) | 5.29 |
| 4 | Gemini 3 Pro (Preview) | 5.03 |
| 5 | O1 | 4.14 |
| 6 | Gemini 2.5 Flash | 3.99 |
| 7 | Gemini 2.5 Pro | 3.82 |
| 8 | Grok 4.1 Fast (Reasoning) | 3.78 |
| 9 | Gemini 2.5 Pro (Preview 03-25) | 3.72 |
| 10 | DeepSeek V3.2 Exp | 3.67 |
| 11 | O3 Mini (High) | 3.64 |
| 12 | Grok 4 | 3.47 |
| 13 | O3 Mini (Medium) | 3.38 |
| 14 | Claude Sonnet 4.5 (Thinking 16K) | 3.37 |
| 15 | Grok 3 Mini (High) | 3.24 |
Interactive version: theaggregate.ai/benchmark?slug=step-game-lechmazur · How the rankings work · Data refreshed daily, snapshot 2026-07-22.