CAPITU - Coherence: leaderboard
Metric: Coherence score (%): a GPT-4o-mini judge's 0-1 rating of the response's coherence and clarity in Brazilian Portuguese given the novel's metadata, scaled to 100, on the 200 single-turn CAPITU prompts (Brazilian Portuguese writing tasks about eight canonical Brazilian novels, each with 1 to 4 verifiable constraints drawn from 59 instruction types such as exact word counts, Portuguese suffix patterns, acrostics and forbidden words), rule-checked on the response as generated, 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 | 99 | #91 |
| 2 | Claude Sonnet 4.5 | 98.8 | #138 |
| 3 | Gemini 3 Pro (Preview) | 98.7 | #64 |
| 4 | GPT-5.2 (Non-reasoning) | 98.4 | #105 (GPT-5.2) |
| 5 | Claude Haiku 4.5 | 98.3 | #271 |
| 6 | GPT-5 Mini | 98.1 | #176 |
| 7 | Gemini 3 Flash (Preview) | 97.7 | #78 |
| 8 | GPT-5.2 (High) | 97.6 | #105 (GPT-5.2) |
| 9 | Qwen 3 235B A22B (Non-reasoning) | 96.8 | #304 (Qwen 3 235B A22B) |
| 10 | Qwen 3 32B (Thinking) | 96.2 | #424 (Qwen 3 32B) |
| 11 | Qwen 3 235B A22B (Thinking) | 96.2 | #304 (Qwen 3 235B A22B) |
| 12 | Qwen 3 14B (Reasoning) | 92.7 | #524 (Qwen 3 14B) |
| 13 | Qwen 3 8B (Thinking) | 92.2 | #667 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=capitu-coherence · How It Works · Data refreshed daily, snapshot 2026-10-11.