AGC-Bench - permpst: leaderboard
The "permpst" subtask of AGC-Bench. A task from AGC-Bench, evaluating models on open-ended creative and figurative-language generation — humor, metaphor, analogy, storytelling and brainstorming — typically graded by LLM-as-judge.
Metric: Dataset z-score. Source: huggingface.co. Saturation forecast: Around June 2027. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 2.14 | #79 |
| 2 | Claude Opus 4.6 | 1.94 | #63 |
| 3 | Claude Sonnet 4.6 | 1.51 | #90 |
| 4 | GLM-5 | 1.35 | #141 |
| 5 | Claude Opus 4.7 | 1.34 | #49 |
| 6 | GPT-5.5 | 1.27 | #31 |
| 7 | Qwen 3 14B | 1.27 | #537 |
| 8 | Hermes-3-Llama-3.1-405B | 1.26 | #510 |
| 9 | Claude Opus 4.5 | 1.18 | #83 |
| 10 | GPT-5.4 Mini | 1.06 | #212 |
| 11 | Kimi K2.6 | 0.96 | #104 |
| 12 | DeepSeek V3 (0324) | 0.91 | #340 |
| 13 | Claude Haiku 4.5 | 0.91 | #280 |
| 14 | Kimi K2.5 | 0.81 | #144 |
| 15 | pixtral-large-2411 | 0.81 | #478 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agc-bench-permpst · How It Works · Data refreshed daily, snapshot 2026-10-20.