Agent Island - Plackett-Luce Strength: leaderboard
Measures model agents in multiplayer games using a shared agent harness. The score is the posterior mean Plackett-Luce strength from the paper tournament, not games played or raw wins. This series reports Plackett-Luce Strength using Posterior Mean Strength. Source: the original paper result table.
Metric: Posterior Mean Strength. Source: arxiv.org. Saturation forecast: Around May 2027. 49 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 5.64 | #33 |
| 2 | GPT-5.2 | 3.1 | #115 |
| 3 | GPT-5.3 Codex | 2.86 | #72 |
| 4 | GPT-5.4 | 2.52 | #83 |
| 5 | Claude Opus 4.7 | 2.07 | #51 |
| 6 | Claude Opus 4.6 | 1.88 | #67 |
| 7 | GPT-5.2 Pro | 1.82 | #70 |
| 8 | DeepSeek V4 Pro | 1.77 | #112 |
| 9 | GLM-5.1 | 1.69 | #109 |
| 10 | Claude Opus 4.5 | 1.64 | #87 |
| 11 | O3 | 1.5 | #132 |
| 12 | Claude Sonnet 4.6 | 1.42 | #94 |
| 13 | GLM-5 | 1.14 | #149 |
| 14 | Gemini 3.1 Pro (Preview) | 1.13 | #64 |
| 15 | Kimi K2 (Thinking) | 1.08 | #252 (Kimi K2) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agent-island-plackett-luce-strength · How It Works · Data refreshed daily, snapshot 2026-11-05.