Mr.LHDR - Web Search VLMs - Final-Answer Accuracy: leaderboard
Mr.LHDR final-answer accuracy (OA) for vision-language models with image input and provider-exposed web search on 102 multimodal long-horizon deep-research questions averaging 12.1 necessary intermediate conclusions (Mr.LHDR, arXiv 2609.11318v1, Table 4; Qwen3-VL-235B judge). Percent; higher is better.
Metric: OA (%; Qwen3-VL-235B judge, 102 items). Source: arxiv.org. Saturation forecast: Around November 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 43.1 | #33 |
| 2 | O4 Mini | 35.3 | #185 |
| 3 | O3 | 34.3 | #132 |
| 4 | GPT-5.4 | 30.4 | #83 |
| 5 | Claude Opus 4.7 | 29.4 | #51 |
| 6 | Grok 4.20 | 22.5 | #179 |
| 7 | GPT-4o (2024-11-20) | 18.6 | #392 |
| 8 | Claude Sonnet 4.6 | 17.6 | #94 |
| 9 | GPT-4.1 | 16.7 | #258 |
| 10 | GPT-5.4 Mini | 14.7 | #228 |
| 11 | GPT-4o Mini (2024-07-18) | 9.8 | #568 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mr-lhdr-web-search-vlms-final-answer-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.