PGG-Bench (Lechmazur): leaderboard
Lech Mazur's PGG-Bench: five LLMs start with 20 tokens and play 10 rounds of public messages, hidden contributions to a fund multiplied by 1.6 and shared, and optional paid punishment of another player. Score is TrueSkill μ on final balances (median over 200 shuffled passes), higher is better. Distinct from the plain contribution-rate LLM Public Goods Game board.
Metric: TrueSkill μ. Source: github.com. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 03-25) | 14.35 | #126 |
| 2 | O1 | 13.4 | #190 |
| 3 | Mistral Large 2 (Jul) | 11.93 | #490 |
| 4 | Gemini 2.0 Pro (Preview 02-05) | 11.14 | #318 |
| 5 | O3 Mini (Medium) | 11.11 | #282 (O3 Mini) |
| 6 | DeepSeek V3 | 10.94 | #329 |
| 7 | GPT-4.5 (Preview) | 10.77 | #260 |
| 8 | Llama 3.3 70B | 10.19 | #578 |
| 9 | Grok 2 (1212) | 10.12 | #502 |
| 10 | DeepSeek R1 | 9.66 | #262 |
| 11 | Qwen 2.5 Max | 9.46 | #427 |
| 12 | Claude 3.7 Sonnet | 9.33 | #259 |
| 13 | Llama 3.1 405B | 8.49 | #476 |
| 14 | Gemini 2.0 Flash (01-21) (Thinking) | 8.26 | #333 (Gemini 2.0 Flash (01-21)) |
| 15 | Claude 3.5 Sonnet (20241022) | 8.05 | #302 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pgg-bench-lechmazur · How It Works · Data refreshed daily, snapshot 2026-11-05.