VistaHop: leaderboard
VistaHop (2026) is a Visual DeepSearch benchmark of 600 long-horizon questions about images that require repeated inspection of image regions and multi-hop web evidence. Models run in the VistaArena agent loop with text search, reverse image search and image cropping, and a judge checks the first answer; the score is Pass@1 in percent, higher is better.
Metric: Pass@1 (%) on 600 Visual DeepSearch tasks (25 scenarios; 154 with 5 to 9 evidence steps and 446 with 10 or more) that require repeatedly inspecting image regions and chaining visual and web evidence, run in the VistaArena agent loop (at most 10 rounds, one tool call per round, temperature 0.7, five runs averaged) with a Qwen3-VL-32B-Instruct judge, with text search, reverse image search and image cropping; higher is better. Source: arxiv.org. Saturation forecast: Around February 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 25.83 | #176 |
| 2 | Claude Sonnet 4.5 | 23.68 | #250 |
| 3 | Gemini 2.5 Pro | 21.83 | #266 |
| 4 | Qwen 3 VL 235B A22B Instruct | 20.54 | #522 |
| 5 | Qwen 3 VL 30B A3B Instruct | 18.34 | #871 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vistahop · How It Works · Data refreshed daily, snapshot 2026-11-13.