ATP-Bench - Manual: leaderboard
Metric: Final Score (0-100) assigned by the Chief Judge of a Gemini 2.5 Pro multi-agent judging system to the model's interleaved text-and-tool-call answer, on the 1,025 Manual queries; zero-shot, tool calls planned but not executed; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 80.58 | #77 |
| 2 | Claude Sonnet 4 | 63.37 | #194 |
| 3 | Claude Sonnet 4.5 | 62.74 | #138 |
| 4 | GPT-5 | 61 | #91 |
| 5 | GPT-4o | 56.38 | #333 |
| 6 | Grok 4.1 Fast (Reasoning) | 55.39 | #208 (Grok 4.1 Fast) |
| 7 | Qwen 2.5 VL 72B Instruct | 41.9 | #364 |
| 8 | Llama 3.2 11B | 27.34 | #1183 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=atp-bench-manual · How It Works · Data refreshed daily, snapshot 2026-10-11.