TRACE Bench - Overall: leaderboard
Weighted roleplay score from agentic checklist evaluation: an LLM user agent elicits role requirements over 200 cases and 5,498 checklist items, combining 0.45 character consistency, 0.05 short-term memory, 0.10 diversity, 0.25 language quality and 0.15 length (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent; higher is better; thinking disabled for all targets; the top is saturated.
Metric: Weighted overall score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (Non-reasoning) | 96.02 | #212 (Gemma 4 31B) |
| 2 | GPT-5.5 | 95.3 | #33 |
| 3 | Claude Sonnet 4.6 | 94.72 | #94 |
| 4 | MiMo-V2.5 | 94.43 | #196 |
| 5 | Kimi K2.5 (Non-reasoning) | 94.31 | #152 (Kimi K2.5) |
| 6 | MiMo-V2.5-Pro | 93.98 | #167 |
| 7 | DeepSeek V4 Pro (Non-reasoning) | 93.96 | #112 (DeepSeek V4 Pro) |
| 8 | MiniMax-M2.7 | 93.88 | #270 |
| 9 | DeepSeek V4 Flash | 92.79 | #155 |
| 10 | GLM-5 (Non-reasoning) | 92.28 | #149 (GLM-5) |
| 11 | Qwen 3.5 27B (Non-reasoning) | 92.1 | #192 (Qwen 3.5 27B) |
| 12 | Qwen 3 14B (Non-reasoning) | 90.62 | #542 (Qwen 3 14B) |
| 13 | Qwen 3 8B (Non-reasoning) | 88.43 | #678 (Qwen 3 8B) |
| 14 | Mistral Small 3.2 | 85.71 | #608 |
| 15 | Ministral 3 14B | 78.55 | #664 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trace-bench-overall · How It Works · Data refreshed daily, snapshot 2026-11-05.