AgentVQA - Web Agents: leaderboard
AgentVQA:offline multiple-choice visual questions sampled across five agentic domains,200questions/domain. Measures decisions from observations,not online task execution. Explicit High/None GPT settings and GLM4.1V Base/Thinking remain distinct; no provider-default effort substitution. Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2030. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (High) | 47 | #20 (GPT-5.6 Sol) |
| 2 | Qwen 3.7 Flash | 43 | #140 |
| 3 | Gemini 3.7 Flash | 40.5 | #22 |
| 4 | Claude Sonnet 5 | 38 | #60 |
| 5 | Claude Opus 4.8 | 35.5 | #36 |
| 6 | Gemini 3.1 Pro (Preview) | 34.5 | #59 |
| 7 | Qwen 3.5 35B A3B | 34.5 | #259 |
| 8 | GPT-5.6 Luna (High) | 34.5 | #73 (GPT-5.6 Luna) |
| 9 | Qwen 3.5 27B | 32.5 | #190 |
| 10 | Qwen 3.7 Plus | 32.5 | #98 |
| 11 | Qwen 3.8 27B | 30 | #111 |
| 12 | Qwen 3.5 122B A10B | 26 | #177 |
| 13 | GLM-4.6V | 25.5 | #317 |
| 14 | Qwen 3.5 397B A17B | 25 | #148 |
| 15 | GLM-4.5V | 25 | #348 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agentvqa-web-agents · How It Works · Data refreshed daily, snapshot 2026-10-24.