PodBench - Script Quality: leaderboard
PodBench's second judging stage: Claude-4.5-Opus scores each podcast script on a 100-point audio-first rubric, 45 points for content depth and insight, 30 for structure and narrative engagement and 25 for conversational language, quoting the script as evidence; higher is better. Proprietary models and open models with thinking disabled; Tencent, ACL 2026.
Metric: Podcast Script Quality (0-100 rubric: depth 45, structure 30, language 25). Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 Instant | 69.64 | #437 |
| 2 | Gemini 3 Pro (Preview) | 63.85 | #111 |
| 3 | Claude Sonnet 4.5 | 63.25 | #244 |
| 4 | DeepSeek V3 (0324) | 59.61 | #704 |
| 5 | Qwen 3 235B A22B (Non-reasoning) | 58.39 | #622 (Qwen 3 235B A22B) |
| 6 | GPT-4o | 56.86 | #741 |
| 7 | Qwen 3 32B (Non-reasoning) | 56.39 | #960 (Qwen 3 32B) |
| 8 | Qwen 3 30B A3B (Non-reasoning) | 52.28 | #1124 (Qwen 3 30B A3B) |
| 9 | Qwen 3 14B (Non-reasoning) | 52.22 | #1235 (Qwen 3 14B) |
| 10 | Qwen 3 8B (Non-reasoning) | 49.93 | #1838 (Qwen 3 8B) |
| 11 | Qwen 2.5 72B Instruct | 49.06 | #1093 |
| 12 | Qwen 3 4B (Non-reasoning) | 47.69 | #2297 (Qwen 3 4B) |
| 13 | Qwen 2.5 32B Instruct | 46.85 | #1216 |
| 14 | InternLM3-8B-Instruct | 45.89 | #2586 |
| 15 | Llama 4 Scout Instruct | 44.51 | #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-script-quality · How It Works · Data refreshed daily, snapshot 2026-11-12.