CCR-Bench - Content-Format Hard Satisfaction: leaderboard
Metric: Hard satisfaction rate (0-1, times 100): share of instructions whose constraints are all satisfied, on CCR-Bench's 64 complex content-format instructions (2-6 tightly coupled content and format constraints each, built on high-difficulty instructions by Gemini-2.5-Pro and refined by experts), temperature 0, model-plus-rule scoring averaged over 10 assessments; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 16.6 | #266 |
| 2 | DeepSeek R1 0528 | 15.8 | #217 |
| 3 | QwQ-32B | 12.2 | #410 |
| 4 | GPT-4.1 | 11.9 | #240 |
| 5 | DeepSeek V3 (0324) | 9.7 | #332 |
| 6 | Qwen 3 32B (Thinking) | 9.4 | #424 (Qwen 3 32B) |
| 7 | Qwen 2.5 72B Instruct | 9.1 | #436 |
| 8 | Qwen 3 32B (Non-reasoning) | 8.9 | #424 (Qwen 3 32B) |
| 9 | Gemini 2.5 Pro | 6.4 | #145 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=ccr-bench-content-format-hard-satisfaction · How It Works · Data refreshed daily, snapshot 2026-10-11.