Pix2Fact (Web Search): leaderboard
Metric: Accuracy (%) on the 1,000 Pix2Fact questions (one per 4K-to-8K real-world photo across eight urban scene categories; each question needs a pixel-level visual clue and external knowledge, written and reviewed by PhD annotators), answers judged against the reference by a Gemini-3.1-Pro LLM judge; temperature 0 (GPT-5.4 at its fixed temperature 1), images resized to API or 15-megapixel limits where needed, and an API call failing three times counts as wrong; original full image, web search allowed: proprietary models through their providers' native search tools, open-weight models through the authors' ReAct agent with a You.com search tool and a Jina Reader page-visit tool (128K context); Claude Opus 4.7 is not run with search (condition C2); higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 42.4 | #54 |
| 2 | Gemini 2.5 Pro | 29.1 | #145 |
| 3 | Grok 4.20 | 22.3 | #167 |
| 4 | GPT-5.4 | 17.9 | #76 |
| 5 | Qwen 3.6 27B | 16.8 | #163 |
| 6 | GLM-4.6V | 11.4 | #309 |
| 7 | Seed 2.0 Pro | 8.6 | #111 |
| 8 | Gemma 4 31B (IT) | 8.2 | #171 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=pix2fact-web-search · How It Works · Data refreshed daily, snapshot 2026-10-11.