{"benchmark_id":"osworld-verified","name":"OSWorld-Verified","parent_benchmark":{"id":"osworld","name":"OSWorld"},"categories":["multimodal","general","agents","vision"],"modality":"multimodal","multilingual":false,"max_score":1.0,"language":"en","description":"OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.","paper_link":"https://arxiv.org/abs/2404.07972","implementation_link":null,"verified":false,"created_at":"2026-05-07T16:53:25.675244+00:00","updated_at":"2026-07-17T06:43:07.876725+00:00","statistics":{"total_models":18,"average_score":0.7015333333333333,"min_score":0.39,"max_score":0.85,"score_stddev":0.12131462256075652,"verified_count":0,"self_reported_count":18},"child_benchmarks":[],"linked_dataset":null,"models":[{"rank":1,"model_id":"claude-fable-5","model_name":"Claude Fable 5","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.85,"normalized_score":0.85,"verified":false,"self_reported":true,"self_reported_source":"https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf","analysis_method":"Fable 5 production-safeguarded configuration. OSWorld score reflects the updated harness with zoom-tool bug fix and 128K max tokens per turn.","verification_date":null,"provider_id":"anthropic","input_cost_per_million":10.0,"output_cost_per_million":50.0,"context_window":1000000,"announcement_date":"2026-06-09","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Anthropic","best_throughput":null,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":2,"model_id":"claude-opus-4-8","model_name":"Claude Opus 4.8","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.834,"normalized_score":0.834,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-opus-4-8","analysis_method":"Agentic computer use evaluation. 361 tasks at 100 steps, 1080p resolution, adaptive thinking at max effort. Pass@1 averaged over 5 seeds. Anthropic updated harness methodology (zoom-tool fix, 128K max tokens per turn).","verification_date":null,"provider_id":"anthropic","input_cost_per_million":5.0,"output_cost_per_million":25.0,"context_window":1000000,"announcement_date":"2026-05-28","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.5,"latency_provider":"Anthropic","best_throughput":42.0,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":3,"model_id":"claude-sonnet-5","model_name":"Claude Sonnet 5","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.812,"normalized_score":0.812,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-sonnet-5","analysis_method":"Agentic computer use. 361 tasks at 100 steps, 1080p resolution, adaptive thinking at max effort. First-attempt success rate averaged over 5 runs. Updated harness (zoom-tool bug fix, 128K max tokens per turn).","verification_date":null,"provider_id":"anthropic","input_cost_per_million":3.0,"output_cost_per_million":15.0,"context_window":1000000,"announcement_date":"2026-06-30","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.4,"latency_provider":"Anthropic","best_throughput":42.0,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":4,"model_id":"claude-mythos-preview","model_name":"Claude Mythos Preview","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.796,"normalized_score":0.796,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/project-glasswing","analysis_method":null,"verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2026-04-07","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":5,"model_id":"gpt-5.5","model_name":"GPT-5.5","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.787,"normalized_score":0.787,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-5/","analysis_method":"OSWorld-Verified desktop navigation benchmark. Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":5.0,"output_cost_per_million":30.0,"context_window":1050000,"announcement_date":"2026-04-23","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"OpenAI","best_throughput":null,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":6,"model_id":"gemini-3.5-flash","model_name":"Gemini 3.5 Flash","organization_id":"google","organization_name":"Google","organization_country":"US","score":0.784,"normalized_score":0.784,"verified":false,"self_reported":true,"self_reported_source":"https://deepmind.google/models/evals-methodology/gemini-3-5-flash/","analysis_method":null,"verification_date":null,"provider_id":"google","input_cost_per_million":1.5,"output_cost_per_million":9.0,"context_window":1048576,"announcement_date":"2026-05-19","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Google","best_throughput":null,"throughput_provider":"Google","context_provider":"Google"},{"rank":7,"model_id":"claude-opus-4-7","model_name":"Claude Opus 4.7","organization_id":"anthropic","organization_name":"Anthropic","organization_country":"US","score":0.78,"normalized_score":0.78,"verified":false,"self_reported":true,"self_reported_source":"https://www.anthropic.com/news/claude-opus-4-7","analysis_method":"Agentic computer use evaluation.","verification_date":null,"provider_id":"anthropic","input_cost_per_million":5.0,"output_cost_per_million":25.0,"context_window":1000000,"announcement_date":"2026-04-16","param_count":null,"is_open_source":false,"is_new":false,"best_latency":0.5,"latency_provider":"Anthropic","best_throughput":42.0,"throughput_provider":"Anthropic","context_provider":"Anthropic"},{"rank":8,"model_id":"gpt-5.4","model_name":"GPT-5.4","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.75,"normalized_score":0.75,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-4/","analysis_method":"OSWorld-Verified desktop navigation benchmark. Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":2.5,"output_cost_per_million":15.0,"context_window":1000000,"announcement_date":"2026-03-05","param_count":null,"is_open_source":false,"is_new":false,"best_latency":3.0,"latency_provider":"OpenAI","best_throughput":50.0,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":9,"model_id":"qwen3.7-plus","model_name":"Qwen3.7-Plus","organization_id":"qwen","organization_name":"Alibaba Cloud / Qwen Team","organization_country":"CN","score":0.733,"normalized_score":0.733,"verified":false,"self_reported":true,"self_reported_source":"https://qwen.ai/blog?id=qwen3.7-plus","analysis_method":"enable_thinking=False","verification_date":null,"provider_id":"together","input_cost_per_million":0.32,"output_cost_per_million":1.28,"context_window":1000000,"announcement_date":"2026-05-31","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Together","best_throughput":null,"throughput_provider":"Together","context_provider":"Together"},{"rank":10,"model_id":"kimi-k2.6","model_name":"Kimi K2.6","organization_id":"moonshotai","organization_name":"Moonshot AI","organization_country":"CN","score":0.731,"normalized_score":0.731,"verified":false,"self_reported":true,"self_reported_source":"https://www.kimi.com/blog/kimi-k2-6","analysis_method":null,"verification_date":null,"provider_id":"deepinfra","input_cost_per_million":0.75,"output_cost_per_million":3.5,"context_window":262144,"announcement_date":"2026-04-20","param_count":1000000000000,"is_open_source":true,"is_new":false,"best_latency":null,"latency_provider":"DeepInfra","best_throughput":null,"throughput_provider":"DeepInfra","context_provider":"Fireworks"},{"rank":11,"model_id":"gpt-5.4-mini","model_name":"GPT-5.4 mini","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.721,"normalized_score":0.721,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-4-mini-and-nano/","analysis_method":"OSWorld-Verified desktop navigation benchmark. Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":0.75,"output_cost_per_million":4.5,"context_window":400000,"announcement_date":"2026-03-17","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"OpenAI","best_throughput":null,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":12,"model_id":"minimax-m3","model_name":"MiniMax M3","organization_id":"minimax","organization_name":"MiniMax","organization_country":"CN","score":0.7006,"normalized_score":0.7006,"verified":false,"self_reported":true,"self_reported_source":"https://www.minimax.io/blog/minimax-m3","analysis_method":"Max Steps = 200","verification_date":null,"provider_id":"fireworks","input_cost_per_million":0.3,"output_cost_per_million":1.2,"context_window":1000000,"announcement_date":"2026-06-01","param_count":null,"is_open_source":true,"is_new":false,"best_latency":null,"latency_provider":"Fireworks","best_throughput":null,"throughput_provider":"Fireworks","context_provider":"Novita"},{"rank":13,"model_id":"gpt-5.3-codex","model_name":"GPT-5.3 Codex","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.647,"normalized_score":0.647,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-3-codex/","analysis_method":"xhigh effort.","verification_date":null,"provider_id":"openai","input_cost_per_million":1.75,"output_cost_per_million":14.0,"context_window":400000,"announcement_date":"2026-02-05","param_count":null,"is_open_source":false,"is_new":false,"best_latency":3.0,"latency_provider":"OpenAI","best_throughput":50.0,"throughput_provider":"OpenAI","context_provider":"OpenAI"},{"rank":14,"model_id":"qwen3.6-plus","model_name":"Qwen3.6 Plus","organization_id":"qwen","organization_name":"Alibaba Cloud / Qwen Team","organization_country":"CN","score":0.625,"normalized_score":0.625,"verified":false,"self_reported":true,"self_reported_source":"https://qwen.ai/blog?id=qwen3.6","analysis_method":null,"verification_date":null,"provider_id":"together","input_cost_per_million":0.5,"output_cost_per_million":3.0,"context_window":1000000,"announcement_date":"2026-04-02","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"Together","best_throughput":null,"throughput_provider":"Together","context_provider":"Together"},{"rank":15,"model_id":"qwen3.5-122b-a10b","model_name":"Qwen3.5-122B-A10B","organization_id":"qwen","organization_name":"Alibaba Cloud / Qwen Team","organization_country":"CN","score":0.58,"normalized_score":0.58,"verified":false,"self_reported":true,"self_reported_source":"https://qwen.ai/blog?id=qwen3.5","analysis_method":null,"verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2026-02-24","param_count":122000000000,"is_open_source":true,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":16,"model_id":"qwen3.5-27b","model_name":"Qwen3.5-27B","organization_id":"qwen","organization_name":"Alibaba Cloud / Qwen Team","organization_country":"CN","score":0.562,"normalized_score":0.562,"verified":false,"self_reported":true,"self_reported_source":"https://qwen.ai/blog?id=qwen3.5","analysis_method":null,"verification_date":null,"provider_id":"novita","input_cost_per_million":0.3,"output_cost_per_million":2.4,"context_window":262144,"announcement_date":"2026-02-24","param_count":27000000000,"is_open_source":true,"is_new":false,"best_latency":null,"latency_provider":"Novita","best_throughput":null,"throughput_provider":"Novita","context_provider":"Novita"},{"rank":17,"model_id":"qwen3.5-35b-a3b","model_name":"Qwen3.5-35B-A3B","organization_id":"qwen","organization_name":"Alibaba Cloud / Qwen Team","organization_country":"CN","score":0.545,"normalized_score":0.545,"verified":false,"self_reported":true,"self_reported_source":"https://qwen.ai/blog?id=qwen3.5","analysis_method":null,"verification_date":null,"provider_id":null,"input_cost_per_million":null,"output_cost_per_million":null,"context_window":null,"announcement_date":"2026-02-24","param_count":35000000000,"is_open_source":true,"is_new":false,"best_latency":null,"latency_provider":null,"best_throughput":null,"throughput_provider":null,"context_provider":null},{"rank":18,"model_id":"gpt-5.4-nano","model_name":"GPT-5.4 nano","organization_id":"openai","organization_name":"OpenAI","organization_country":"US","score":0.39,"normalized_score":0.39,"verified":false,"self_reported":true,"self_reported_source":"https://openai.com/index/introducing-gpt-5-4-mini-and-nano/","analysis_method":"OSWorld-Verified desktop navigation benchmark. Reasoning effort xhigh.","verification_date":null,"provider_id":"openai","input_cost_per_million":0.2,"output_cost_per_million":1.25,"context_window":400000,"announcement_date":"2026-03-17","param_count":null,"is_open_source":false,"is_new":false,"best_latency":null,"latency_provider":"OpenAI","best_throughput":null,"throughput_provider":"OpenAI","context_provider":"OpenAI"}]}