multimodal benchmark
OSWorld Leaderboard
OSWorld is a multimodal benchmark with 369 computer tasks involving real web and desktop applications, OS file I/O, and multi-application workflows across Ubuntu, Windows, and macOS.
Updated Sep 5, 2026
Models20
Model coverage20
MetricScore
EvidenceB
OSWorld Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelBY | Score78.80% | Percentile100.00% | Participants20 | EvidenceC | Evaluated |
| Rank02 | ModelBY | Score76.40% | Percentile94.74% | Participants20 | EvidenceC | Evaluated |
| Rank03 | ModelAN | Score72.70% | Percentile89.47% | Participants20 | EvidenceC | Evaluated |
| Rank04 | ModelAN | Score72.50% | Percentile84.21% | Participants20 | EvidenceC | Evaluated |
| Rank05 | ModelAC | Score66.70% | Percentile78.95% | Participants20 | EvidenceC | Evaluated |
| Rank06 | ModelAN | Score66.30% | Percentile73.68% | Participants20 | EvidenceC | Evaluated |
| Rank07 | ModelZA | Score62.30% | Percentile68.42% | Participants20 | EvidenceC | Evaluated |
| Rank08 | ModelAN | Score61.40% | Percentile63.16% | Participants20 | EvidenceC | Evaluated |
| Rank09 | ModelAN | Score50.70% | Percentile57.89% | Participants20 | EvidenceC | Evaluated |
| Rank10 | ModelAC | Score41.00% | Percentile52.63% | Participants20 | EvidenceC | Evaluated |
| Rank11 | ModelAC | Score38.10% | Percentile47.37% | Participants20 | EvidenceC | Evaluated |
| Rank12 | ModelAC | Score33.90% | Percentile42.11% | Participants20 | EvidenceC | Evaluated |
| Rank13 | ModelAC | Score33.90% | Percentile36.84% | Participants20 | EvidenceC | Evaluated |
| Rank14 | ModelAC | Score32.60% | Percentile31.58% | Participants20 | EvidenceC | Evaluated |
| Rank15 | ModelAC | Score31.40% | Percentile26.32% | Participants20 | EvidenceC | Evaluated |
| Rank16 | ModelAC | Score30.60% | Percentile21.05% | Participants20 | EvidenceC | Evaluated |
| Rank17 | ModelAC | Score30.30% | Percentile15.79% | Participants20 | EvidenceC | Evaluated |
| Rank18 | ModelAC | Score26.20% | Percentile10.53% | Participants20 | EvidenceC | Evaluated |
| Rank19 | ModelAC | Score8.83% | Percentile5.26% | Participants20 | EvidenceC | Evaluated |
| Rank20 | ModelAC | Score5.92% | Percentile0.00% | Participants20 | EvidenceC | Evaluated |
OSWorld Highlights
The leading models and scores on this benchmark.
OSWorld Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for OSWorld
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What is OSWorld?
What OSWorld measures and how its scores work.
OSWorld is a scalable real computer environment for multimodal agents that supports task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.
It uses the Score metric, reported as a ratio, for tasks involving real web and desktop applications, OS file I/O, and multi-application workflows.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- OSWorld
- Modality
- multimodal
- Primary category
- multimodal
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- overall
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about OSWorld.
Which model scores highest on OSWorld?
Seed 2.1 Pro is currently ranked first with 78.80%.
What are the top three models on OSWorld?
The current leaders are Seed 2.1 Pro (78.80%), Seed 2.1 Turbo (76.40%), and Claude Opus 4.6 (72.70%).
Which OSWorld model has the lowest official input price?
Claude Haiku 4.5 has the lowest matched official input price at $1.0 input / $5.0 output per 1M tokens.
Which models are fastest among OSWorld results?
The fastest matched records are Claude Opus 4.6 (123.15 tok/s via Anthropic), Claude Haiku 4.5 (100.00 tok/s via Vertex AI), and Claude Opus 4.5 (42.00 tok/s via Anthropic).
Does the highest score result prove overall model quality?
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.