DiG-bench - Games Beaten: leaderboard
DiG-bench (Discovery in Games): 70 text games in 7 difficulty tiers whose rules are undisclosed and must be discovered by experimentation, all beaten by humans on the first attempt. Expected number of games beaten out of 70 (mean win rate over runs × games) with the basic harness; higher is better. Reasoning effort follows the leaderboard's run configuration.
Metric: Games beaten (of 70). Source: digbench.ai. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (High) | 50 | #16 (Claude Opus 5) |
| 2 | GPT-5.5 (High) | 18 | #33 (GPT-5.5) |
| 3 | Kimi K3 (Max) | 16 | #35 (Kimi K3) |
| 4 | Gemini 3.1 Pro (Preview) (High) | 11.75 | #64 (Gemini 3.1 Pro (Preview)) |
| 5 | GLM-5.2 (Max) | 10.67 | #80 (GLM-5.2) |
| 6 | DeepSeek V4 Flash (0731) | 8 | #84 |
| 7 | DeepSeek V4 Pro Preview | 3 | #150 |
| 8 | Qwen 3.6 27B | 1 | #182 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=dig-bench-games-beaten · How It Works · Data refreshed daily, snapshot 2026-11-05.