ScreenSpot-Pro (with tools)
Multimodal · 2024-12-19
ScreenSpot-Pro's 1,581 GUI-grounding instructions answered with tools that act on the screenshot, such as a Python interpreter, an image-cropping or zoom tool, or a capture-screenshot function call. Tool runs refine the predicted point over several passes and score far above single-pass grounding, so they are kept apart; runs without tools are scored under ScreenSpot-Pro.
Benchmark source: arXiv
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| Opus 4.8 | 87.9 | Claude Opus 4.8 system card · Anthropic |
| GPT-5.2 | 86.3 | Gemini 3 Flash model card · Google |
| Muse Spark | 84.1 | Muse Spark launch post and evaluation methodology · Meta |
| Qwen3.6 27B | 76.1 | Muse Glimmer report Table 2, Qwen3.6-27B (thinking) · Meta |
| Gemma 4 31B IT | 75.9 | Muse Glimmer report Table 2, Gemma4-31B (thinking) · Meta |
| Muse Glimmer 30B | 75.7 | Muse Glimmer report Table 5, Muse Glimmer (xhigh) · Meta |
| Gemini 3 Pro Preview | 72.7 | Gemini 3 Flash model card · Google |
| Gemini 3 Flash Preview | 69.1 | Gemini 3 Flash model card · Google |
| GPT-5.1 | 64.2 | GPT-5.2 launch post · OpenAI |