Benchmark

OSWorld Benchmark

What is OSWorld?

OSWorld: The first-of-its-kind scalable, real computer environment for multimodal agents, supporting task setup, execution-based evaluation, and interactive…

Released
2024-04-11
Modality
multimodal
Evaluates
multimodal, general, agents, vision
Openness
unknown
Importer
LLM Stats
Reported scores
20

Source provenance

OSWorld code

OSWorld results and reported scores

LLM Stats

ModelOrganizationReported valueReportedEvidence
Claude Haiku 4.5Anthropic0.5072025-10-15evidence
Claude Opus 4.5Anthropic0.6632025-11-24evidence
Claude Opus 4.6Anthropic0.7272026-02-05evidence
Claude Sonnet 4.5Anthropic0.6142025-09-29evidence
Claude Sonnet 4.6Anthropic0.7252026-02-17evidence
GLM-5V-TurboZ.ai0.6232026-04-02evidence
Qwen2.5 VL 32B InstructQwen0.05922025-02-28evidence
Qwen2.5 VL 72B InstructQwen0.08832025-01-26evidence
Qwen3 VL 235B A22B InstructQwen0.6672025-09-22evidence
Qwen3 VL 235B A22B ThinkingQwen0.3812025-09-22evidence
Qwen3 VL 30B A3B InstructQwen0.3032025-09-22evidence
Qwen3 VL 30B A3B ThinkingQwen0.3062025-09-22evidence
Qwen3 VL 32B InstructQwen0.3262025-09-22evidence
Qwen3 VL 32B ThinkingQwen0.412025-09-22evidence
Qwen3 VL 4B InstructQwen0.2622025-09-22evidence
Qwen3 VL 4B ThinkingQwen0.3142025-09-22evidence
Qwen3 VL 8B InstructQwen0.3392025-09-22evidence
Qwen3 VL 8B ThinkingQwen0.3392025-09-22evidence
Seed 2.1 ProByteDance0.7882026-06-24evidence
Seed 2.1 TurboByteDance0.7642026-06-24evidence

Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.

Which sources cite OSWorld?

Related multimodal, general, agents, vision benchmarks