Inference
Kaixin Li et al.

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.

ModelsModels with a result.18
Top resultBest result on this test.92.7%GPT-6 Astra
Top-3 spreadPoints from first to third.7.3 pts
YearYear of release.2025

Result and price

020406080100$1$3$10$30$100RESULTOUTPUT PRICE PER 1M TOKENS · LOG SCALE

Results

18 results
#ModelResultThe score from the source.
01GPT-6 Astra92.7%
02Claude Opus 4.887.9%
03GPT-5.485.4%
04Qwen3.8 Max84.5%
05Gemini 3.1 Pro84.4%
06Muse Spark84.1%
07Claude Opus 4.683.1%
08Qwen3.7 Plus79.0%
09Muse Glimmer 30B75.4%
10Gemini 3 Pro72.7%
11Holo2-235B-A22B70.6%
12Qwen3.6 Plus68.2%
13Holo2-30B-A3B66.1%
14Qwen3.5 397B65.6%
15Holo2-8B58.9%
16Nemotron 3 Nano Omni 30B A3B57.8%
17Holo2-4B57.2%
18Claude Opus 4.545.7%