PodBench: leaderboard
PodBench (Tencent, ACL 2026) asks a model to write a podcast script from source documents of up to 21K tokens and a user instruction, over 800 Chinese and English queries. Claude-4.5-Opus judges each script for instruction following, against a checklist it derives from the instruction, and for audio-first script quality on a 100-point rubric; the score is their mean, 0-100, higher is better. This board has the proprietary models and the open models run with thinking disabled; runs with thinking enabled are on PodBench - Thinking Mode.
Metric: PodBench Score (0-100; mean of instruction following and script quality). Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 Instant | 82.58 | #437 |
| 2 | Gemini 3 Pro (Preview) | 79.97 | #111 |
| 3 | Claude Sonnet 4.5 | 79.91 | #244 |
| 4 | DeepSeek V3 (0324) | 76.16 | #704 |
| 5 | Qwen 3 235B A22B (Non-reasoning) | 76.02 | #622 (Qwen 3 235B A22B) |
| 6 | Qwen 3 32B (Non-reasoning) | 73.54 | #960 (Qwen 3 32B) |
| 7 | GPT-4o | 73.15 | #741 |
| 8 | Qwen 3 14B (Non-reasoning) | 69.65 | #1235 (Qwen 3 14B) |
| 9 | Qwen 3 30B A3B (Non-reasoning) | 69.39 | #1124 (Qwen 3 30B A3B) |
| 10 | Qwen 2.5 72B Instruct | 66.19 | #1093 |
| 11 | Qwen 3 8B (Non-reasoning) | 65.36 | #1838 (Qwen 3 8B) |
| 12 | Qwen 2.5 32B Instruct | 62.35 | #1216 |
| 13 | Qwen 3 4B (Non-reasoning) | 62.21 | #2297 (Qwen 3 4B) |
| 14 | InternLM3-8B-Instruct | 59.92 | #2586 |
| 15 | Llama 4 Scout Instruct | 57.33 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.