Skip to content

ScreenSpot Pro

A GUI-grounding benchmark for 1,581 instructions in full-screen, high-resolution professional interfaces. It tests where a target is, not whether an agent can finish the surrounding workflow.

15 models

Higher is better.

  1. 1GPT-6 AstraOpenAI92.7
  2. 2Claude Opus 4.8Anthropic87.9
  3. 3GPT-5.4OpenAI85.4
  4. 4Qwen3.8 MaxAlibaba84.5
  5. 5Gemini 3.1 ProGoogle84.4
  6. 6Muse SparkMeta84.1
  7. 7Claude Opus 4.6Anthropic83.1
  8. 8Qwen3.7 PlusAlibaba79.0
  9. 9Muse Glimmer 30BMeta75.4
  10. 10Gemini 3 ProGoogle72.7
  11. 11Holo2-235B-A22BH Company70.6
  12. 12Qwen3.6 PlusAlibaba68.2
  13. 13Qwen3.5 397BAlibaba65.6
  14. 14Nemotron 3 Nano Omni 30B A3BNVIDIA57.8
  15. 15Claude Opus 4.5Anthropic45.7
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Static interface element localization
Tasks
1,581 grounding instructions
Authors
Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, Tat-Seng Chua (2025)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.