General Agents Benchmark

OSWorld-G

OSWorld-G - scaled computer-use grounding benchmark via user interface decomposition and synthesis (NeurIPS 2025 Spotlight). Scores are measured on OSWorld-G tasks; repo: https://github.com/xlang-ai/OSWorld-G (paper: https://arxiv.org/abs/2505.13227).

Category

General Agents

Score type

percent

Scale

100%

higher is better

Models tested

3

What is OSWorld-G?

OSWorld-G is a benchmark for how well a model acts as a general-purpose agent. The official description says: "OSWorld-G - scaled computer-use grounding benchmark via user interface decomposition and synthesis (NeurIPS 2025 Spotlight)". Below we describe in simple words what benchmarks of this type generally examine.

What does it test?

OSWorld-G measures how well a model acts as a general-purpose agent. In detail, it checks:

  • solving long, multi-step tasks with many intermediate decisions
  • using tools correctly (browsers, terminals, APIs, calendars, other apps)
  • adapting when something unexpected happens mid-task
  • staying on goal over long task horizons instead of wandering off

How to read the score

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: OSWorld-G - scaled computer-use grounding benchmark via user interface decomposition and synthesis (NeurIPS 2025 Spotlight). Scores are measured on OSWorld-G tasks; repo: https://github.com/xlang-ai/OSWorld-G (paper: https://arxiv.org/abs/2505.13227).

Frequently asked questions

What is the OSWorld-G benchmark?

OSWorld-G is a benchmark for how well a model acts as a general-purpose agent. The official description says: 'OSWorld-G - scaled computer-use grounding benchmark via user interface decomposition and synthesis (NeurIPS 2025 Spotlight)'. Below we describe in simple words what benchmarks of this type generally examine.

What does OSWorld-G test?

OSWorld-G mainly measures how well a model acts as a general-purpose agent. Concretely, it checks: solving long, multi-step tasks with many intermediate decisions; using tools correctly (browsers, terminals, APIs, calendars, other apps); adapting when something unexpected happens mid-task.

How is the OSWorld-G score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get OSWorld-G data and scores?

The official source is https://github.com/xlang-ai/OSWorld-G. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/osworld-g/scores) or pull it directly from the Open LLM Radar API with any HTTP client.

Model scores 3 models tested

Reset
# Model Score Date Source Verified
1 Sage-S2.1-Flash
Shanghai AI Laboratory
84.8%
— Report Verified
2 GLM-5.3-Flash
Zhipu AI
83.7%
— Report Unverified
3 DeepSeek-V4.1-Flash-1017
DeepSeek
61.4%
— Report Unverified

What now?

You've seen how models score on OSWorld-G — here's what to do next.