CAPITU - Multi-Turn: leaderboard
Metric: Conversation accuracy (%): share of the 100 three-turn CAPITU conversations about Brazilian novels in which every turn satisfies all of its accumulated constraints (one new verifiable constraint per turn), rule-checked strictly, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 96 | #176 |
| 2 | Gemini 3 Pro (Preview) | 95 | #64 |
| 3 | Gemini 3 Flash (Preview) | 95 | #78 |
| 4 | GPT-5 | 90 | #91 |
| 5 | Claude Sonnet 4.5 | 87 | #138 |
| 6 | GPT-5.2 (High) | 83 | #105 (GPT-5.2) |
| 7 | Qwen 3 14B (Reasoning) | 74 | #524 (Qwen 3 14B) |
| 8 | GPT-5.2 (Non-reasoning) | 73 | #105 (GPT-5.2) |
| 9 | Claude Haiku 4.5 | 72 | #271 |
| 10 | Qwen 3 8B (Thinking) | 72 | #667 (Qwen 3 8B) |
| 11 | Qwen 3 235B A22B (Thinking) | 64 | #304 (Qwen 3 235B A22B) |
| 12 | Qwen 3 32B (Thinking) | 60 | #424 (Qwen 3 32B) |
| 13 | Qwen 3 235B A22B (Non-reasoning) | 60 | #304 (Qwen 3 235B A22B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=capitu-multi-turn · How It Works · Data refreshed daily, snapshot 2026-10-11.