Pencil Puzzle Bench - Firefly: leaderboard
Models draw Firefly paths from directional clues, coordinating path lengths and grid constraints without crossings. Direct-ask exact success fraction for 15fixed golden-300 puzzles of this type; archived DirectAskStrategy_05aaaac9, latest record per source model and puzzle. No agentic or revised strategy cohort mixed.
Metric: Direct-ask Success Rate (%). Source: ppbench.com. Saturation forecast: Around March 2027. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (xHigh) | 33.33 | #110 (GPT-5.2) |
| 2 | Claude Opus 4.6 (Thinking) | 20 | #63 (Claude Opus 4.6) |
| 3 | GPT-5.2 (High) | 13.33 | #110 (GPT-5.2) |
| 4 | Gemini 3.1 Pro (Preview) | 6.67 | #58 |
| 5 | GPT-4o | 0 | #343 |
| 6 | Gemini 2.5 Pro | 0 | #152 |
| 7 | Claude Sonnet 4.6 | 0 | #90 |
| 8 | Claude Sonnet 4.5 | 0 | #142 |
| 9 | GPT-5.2 | 0 | #110 |
| 10 | GPT-4.1 | 0 | #248 |
| 11 | Claude Opus 4.6 | 0 | #63 |
| 12 | GPT-OSS-120B | 0 | #341 |
| 13 | Kimi K2.5 | 0 | #144 |
| 14 | DeepSeek V3.2 | 0 | #205 |
| 15 | Qwen 3.5 397B A17B | 0 | #146 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pencil-puzzle-bench-firefly · How It Works · Data refreshed daily, snapshot 2026-10-20.