This reporting page focuses on computer-use and GUI-agent behavior: whether a model can read screens, ground actions, and complete software tasks. It is distinct from pure tool calling and distinct from plain multimodal image understanding.
This page ranks models using only sourced computer-use and GUI benchmarks in the reporting family.
Bottom line: Computer-use AI is still early — only a handful of models have verifiable GUI grounding scores. GPT-5.4 and Claude Opus 4.6 lead on OSWorld-Verified.
According to BenchLM.ai, GPT-5.4 leads this ranking with a score of 79.2, followed by Claude Opus 4.6 (76.9) and Claude Opus 4.5 (58.1). There is a significant gap between the leading models and the rest of the field.
All models in this ranking are proprietary. No open-weight alternatives are available for this category.
This ranking is based on provisional overall weighted scores across BenchLM.ai's scoring formula tracked by BenchLM.ai. For detailed model profiles, click any model name below. To compare two specific models head-to-head, use the "vs #" links.
GPT-5.4
OpenAI · 1.05M
Claude Opus 4.6
Anthropic · 1M
Strong GUI grounding. Reliable for desktop automation tasks.
Claude Opus 4.5
Anthropic · 200K
Previous-gen Anthropic model with good computer-use baseline.
GPT-5.4 leads computer-use with the best OSWorld-Verified and ScreenSpot Pro scores.
Claude Opus 4.6 close second with strong GUI grounding across benchmarks.
Claude Opus 4.5 holds third with solid OSWorld coverage.
Get notified when models move. One email a week with what changed and why.
Free. No spam. Unsubscribe anytime.
The top model on this sourced reporting-family slice is GPT-5.4 by OpenAI with an average of 79.2.
3 models are listed with sourced benchmark coverage in this reporting family.
This is a reporting family ranking, not a weighted category. It averages sourced computer-use and GUI benchmarks to give a focused view of this capability.
Models must have sourced results on at least a quarter of the benchmarks in this family to be included. Coverage varies — a model with 2 benchmark scores is less reliable than one with 5.
For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.
Free. No spam. Unsubscribe anytime.