TRACE Bench - Character Consistency: leaderboard
Share of prebuilt role-profile checklist items (excluding memory probes) that a roleplay model completes during agent-driven dialogues (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent; higher is better. 26 API, open and roleplay-tuned models, temperature 0.8 with thinking disabled.
Metric: Checklist completion rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (Non-reasoning) | 95.46 | #212 (Gemma 4 31B) |
| 2 | MiMo-V2.5-Pro | 94.95 | #167 |
| 3 | Claude Sonnet 4.6 | 94.68 | #94 |
| 4 | Qwen 3.5 27B (Non-reasoning) | 94.53 | #192 (Qwen 3.5 27B) |
| 5 | GPT-5.5 | 94.42 | #33 |
| 6 | Kimi K2.5 (Non-reasoning) | 94.25 | #152 (Kimi K2.5) |
| 7 | MiniMax-M2.7 | 94 | #270 |
| 8 | MiMo-V2.5 | 93.94 | #196 |
| 9 | GLM-5 (Non-reasoning) | 93.94 | #149 (GLM-5) |
| 10 | DeepSeek V4 Pro (Non-reasoning) | 93.56 | #112 (DeepSeek V4 Pro) |
| 11 | DeepSeek V4 Flash | 91.81 | #155 |
| 12 | Qwen 3 14B (Non-reasoning) | 88.36 | #542 (Qwen 3 14B) |
| 13 | Ministral 3 14B | 87.82 | #664 |
| 14 | Mistral Small 3.2 | 87.57 | #608 |
| 15 | Qwen 3 8B (Non-reasoning) | 85.44 | #678 (Qwen 3 8B) |
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-character-consistency · How It Works · Data refreshed daily, snapshot 2026-11-05.