PodBench - Instruction Following: leaderboard
PodBench's first judging stage: for each of 800 Chinese and English podcast-script requests with source documents of up to 21K tokens, Claude-4.5-Opus derives an acceptance checklist of explicit and implied requirements (language, speaker count, focus, length) and scores every item 0, 0.5 or 1, normalized to 0-100, higher is better. Proprietary models and open models with thinking disabled; Tencent, ACL 2026.
Metric: Instruction Following (0-100; judge-derived checklist, items scored 0/0.5/1). Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 96.58 | #244 |
| 2 | Gemini 3 Pro (Preview) | 96.09 | #111 |
| 3 | GPT-5.1 Instant | 95.52 | #437 |
| 4 | Qwen 3 235B A22B (Non-reasoning) | 93.64 | #622 (Qwen 3 235B A22B) |
| 5 | DeepSeek V3 (0324) | 92.71 | #704 |
| 6 | Qwen 3 32B (Non-reasoning) | 90.69 | #960 (Qwen 3 32B) |
| 7 | GPT-4o | 89.43 | #741 |
| 8 | Qwen 3 14B (Non-reasoning) | 87.07 | #1235 (Qwen 3 14B) |
| 9 | Qwen 3 30B A3B (Non-reasoning) | 86.5 | #1124 (Qwen 3 30B A3B) |
| 10 | Qwen 2.5 72B Instruct | 83.31 | #1093 |
| 11 | Qwen 3 8B (Non-reasoning) | 80.79 | #1838 (Qwen 3 8B) |
| 12 | Qwen 2.5 32B Instruct | 77.84 | #1216 |
| 13 | Qwen 3 4B (Non-reasoning) | 76.73 | #2297 (Qwen 3 4B) |
| 14 | InternLM3-8B-Instruct | 73.95 | #2586 |
| 15 | Llama 4 Scout Instruct | 70.14 | #1532 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=podbench-instruction-following · How It Works · Data refreshed daily, snapshot 2026-11-12.