ScreenSpot Pro
We show this table for reference; we do not rank on it.
GPT-6 Astra has the highest reported ScreenSpot Pro score at 92.7%, ahead of Claude Opus 4.8 (87.9%) and GPT-5.4 (85.4%) among 19 models. We compile the table from benchmark-owner or independent runs, provider self-reports and secondary reports and keep it for display only; it does not affect overall rankings.
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.
Benchmark score on ScreenSpot Pro — October 27, 2026
We compile the ScreenSpot Pro rows from benchmark-owner or independent runs, provider self-reports and secondary reports. GPT-6 Astra leads the table at 92.7%, followed by Claude Opus 4.8 (87.9%) and GPT-5.4 (85.4%). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
Claude Opus 4.8
Anthropic
GPT-5.4
OpenAI
19 modelsMultimodal & GroundedCurrentDisplay onlyUpdated October 27, 2026
Benchmark score table (19 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 92.7% | Not reported | Closed |
| 2 | Claude Opus 4.8Anthropic | 87.9% | Not reported | Closed |
| 3 | GPT-5.4OpenAI | 85.4% | Not reported | Closed |
| 4 | Qwen3.8 MaxAlibaba | 84.5% | Not reported | Open |
| 5 | Gemini 3.1 ProGoogle | 84.4% | Not reported | Closed |
| 6 | Muse SparkMeta | 84.1% | Not reported | Closed |
| 7 | Claude Opus 4.6Anthropic | 83.1% | Not reported | Closed |
| 8 | Qwen3.7 PlusAlibaba | 79.0% | Not reported | Closed |
| 9 | Muse Glimmer 30BMeta | 75.4% | Not reported | Open |
| 10 | Gemini 3 ProGoogle | 72.7% | Not reported | Closed |
| 11 | Holo2-235B-A22BH Company | 70.6% | Not reported | Open |
| 12 | Qwen3.6 PlusAlibaba | 68.2% | Not reported | Closed |
| 13 | Holo2-30B-A3BH Company | 66.1% | Not reported | Open |
| 14 | Qwen3.5 397BAlibaba | 65.6% | Not reported | Open |
| 15 | Holo2-8BH Company | 58.9% | Not reported | Open |
| 16 | Nemotron 3 Nano Omni 30B A3BNVIDIA | 57.8% | Not reported | Open |
| 17 | Holo2-4BH Company | 57.2% | Not reported | Open |
| 18 | LFM3-VL-2BLiquidAI | 55.5% | Not reported | Open |
| 19 | Claude Opus 4.5Anthropic | 45.7% | Not reported | Closed |
How to read this leaderboard
Editorial review by Glevd · 2026-07-15
Use ScreenSpot Pro to judge published GUI-grounding systems after checking how each system handles the image. Direct coordinate prediction, cropping, visual search, a planner, or Python tools can change the result without changing the underlying base model.
Operator receipt: 19 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 92.7%.
Honest limit: ScreenSpot Pro stops at localization on a static screenshot. It does not test clicking, typing, state tracking, recovery, or completion of a multi-step workflow, and published rows with different tool or search setups are not a clean model-only comparison.
Among the reported ScreenSpot Pro rows, GPT-6 Astra is first at 92.7%. The third row is 7.3 points behind. The broader top-10 range is 20.0 points, so the table still separates the published systems.
19 models have been evaluated on ScreenSpot Pro. The benchmark falls in the Multimodal & Grounded category. ScreenSpot Pro is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ScreenSpot Pro
Year
2025
Tasks
1,581 grounding instructions
Format
Static interface element localization
Difficulty
Professional GUI grounding
The benchmark spans 23 applications, six application categories, and three operating systems. It scores whether a system localizes the requested text or icon in a static professional screenshot, using micro-average accuracy on the official leaderboard.
Freshness and provenance
Version
ScreenSpot Pro 2025
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ScreenSpot Pro measure?
ScreenSpot Pro measures whether a model can map a natural-language instruction to the correct target in a full-screen, high-resolution professional interface. Its 1,581 examples span 23 applications, six application categories, and three operating systems. The score is grounding accuracy, not end-to-end task completion.
What does a high ScreenSpot Pro score mean?
A high score means the published system localized more requested interface elements under its reported setup. It does not prove the same base model will click correctly inside an agent. Cropping, visual search, planners, Python tools, image resolution, and decoding policy can materially change the result.
Can ScreenSpot Pro pick the best computer-use agent?
No. ScreenSpot Pro tests static-screen grounding, which is one prerequisite for computer use. It does not test state tracking, typing, recovery after a bad action, or completion of a multi-step workflow. Pair it with OSWorld Verified and an application-specific operator test before choosing an agent.
Compare top models on ScreenSpot Pro
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.