LongProc: leaderboard
Long-context procedural generation benchmark testing models on generating long, structured outputs following complex procedural instructions.
Metric: Accuracy @ 0.5K (%). Source: princeton-pli.github.io. Saturation forecast: Estimated already saturated. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 94.8 | #336 |
| 2 | Gemini 1.5 Pro (001) | 89.2 | #444 |
| 3 | Gemini 1.5 Flash (001) | 78.9 | #577 |
| 4 | Claude 3.5 Sonnet (20241022) | 78.4 | #291 |
| 5 | Llama 3.1 70B Instruct | 72.9 | #561 |
| 6 | Qwen 2.5 72B Instruct | 68.7 | #454 |
| 7 | Llama 3.1 8B Instruct | 26.9 | #1035 |
| 8 | Llama 3.2 3B Instruct | 13.5 | #1352 |
| 9 | Llama 3.2 1B Instruct | 4 | #1539 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=longproc · How It Works · Data refreshed daily, snapshot 2026-10-20.