- OSWorld 2.0 (end-to-end models) by OSWorld team, XLANG Lab, The University of Hong Kong (automated tests, Binary accuracy: share of 108 long-horizon computer-use workflows fully completed (default view: 500-step budget, full task set)) · snapshot 15 Oct 2026
Default view (end-to-end models, 500 steps, full task set, all task releases): 23 rows covering 11 model versions, incl. reasoning-effort variants; task releases v2026.06.24, v2026.08.08 and v2.1 differ, so scores across releases are not strictly comparable. Data file static/data/leaderboard/official-results.json has updatedAt 2026-10-15 (rows for MuseSpark 1.3 and 1.2 added then; repository xlang-ai/OSWorld-V2 pushed 2026-10-15); all model rows are marked official (run by the OSWorld team). Lists Claude Opus 5 (leader, a previous version), GPT-5.6 Sol, MuseSpark 1.3, MiniMax M3, Kimi 2.6, Qwen 3.7-Plus, older Claude/GPT; current tracked versions present: Muse Spark 1.3, MiniMax M3, Qwen3.7-Plus. No Gemini entries and no Claude Opus/Sonnet 5.5, Fable, GPT-6.x or Gemini 4 Argon rows; vendor-announced OSWorld 2.0 numbers are not on this board. Re-checked 2026-10-19: data file official-results.json still updatedAt 2026-10-15 (55 rows across views, unchanged); repo xlang-ai/OSWorld-V2 last pushed 2026-10-15 (harness commits only). Still stale for Claude Opus/Sonnet 5.5, Fable 5.1, GPT-6.x, Gemini 4 Argon.
- OSWorld-Verified (models) by OSWorld team, XLANG Lab, The University of Hong Kong (automated tests, Success rate: share of OSWorld's ~360 Ubuntu desktop tasks (browser, office, media, code, multi-app) completed end to end, re-run by the OSWorld team) · snapshot 19 Oct 2026
93 scored rows of approach type 'General model' or 'Specialized model' (42 distinct names; one row per max-steps setting 15/50/100). Rows read from the page's data file static/data/osworld_verified_results.xlsx; no last-updated date shown; newest model entry dated 2026-08-01 (claude-fable-5[1m] 85.96 and claude-opus-5[1m] 83.39, both previous versions) and the site repository's last data commit is 2026-08-07. Frontier versions current on 2026-10-19 (Claude Opus 5.5, Sonnet 5.5, Fable 5.1, GPT-6.x, Gemini 4 Argon, Kimi K3, Holo4, Fara1.5) are not listed. The page carries a notice dated 2026-06-26 that OSWorld 2.0 (osworld-v2.xlang.ai) is the latest version; see osworld-2. Re-checked 2026-10-19: site repository os-world/os-world.github.io last commit still 2026-08-07, so the xlsx is unchanged; still stale.
- OSWorld-Verified (agentic frameworks) by OSWorld team, XLANG Lab, The University of Hong Kong (automated tests, Success rate: share of OSWorld's ~360 Ubuntu desktop tasks completed end to end by an agent framework plus its models, re-run by the OSWorld team, ranks tools) · snapshot 19 Oct 2026
52 scored rows of approach type 'Agentic framework' (24 distinct names, incl. multi-rollout best-of-N runs and max-steps variants). Mostly research and enterprise agents (Intelligence-Indeed Agent, Pointer Agent w/ Opus 4.7, OpenAPA w/ gemini-3.1-pro, VLAA-GUI, HIPPO, UiPath Screen Agent, OS-Symphony); newest entry 2026-07-25; site repository's last data commit 2026-08-07. Same page and data file as osworld-verified; no last-updated date shown; the page points to OSWorld 2.0 as the latest version (osworld-2-tools has only 3 agent rows so far). Re-checked 2026-10-19: unchanged (last repository commit 2026-08-07).
- OSWorld 2.0 (agent systems) by OSWorld team, XLANG Lab, The University of Hong Kong (automated tests, Binary accuracy: share of 108 long-horizon computer-use workflows fully completed by an agent system (harness plus model), ranks tools) · snapshot 15 Oct 2026
3 agent-system rows (shown when the model filter is set to all): Darwinx w/ Claude Opus 4.8 (Salesforce AI Research) at a 450-step budget (44.44 on task release v2.1, 37.65 on v2026.08.08) and Sai w/ Claude Opus 5 (Simular) at 500 steps (41.67, v2.1, 108 tasks). Same page, data file and update date (2026-10-15) as osworld-2; verified unchanged 2026-10-19. Re-checked 2026-10-19: unchanged.
- Online-Mind2Web (auto-eval, WebJudge o4-mini) by OSU NLP Group, The Ohio State University (automated tests, Average success rate on 300 live tasks across 136 real websites (easy/medium/hard), judged automatically by WebJudge (o4-mini), ranks tools) · snapshot 5 Aug 2026
20 rows in the auto-eval CSV (auto_o4-mini_Mind2Web-Online - Leaderboard_data.csv); Space last modified 2026-08-05, newest entries 2026-08-04. Entries are agent systems (Operator, Browser Use w/ GPT-4o, Navigator n1.5, ACT-2, Hark, Webwright, ChatGPT Atlas) plus a few model-in-harness rows (Gemini 2.5 Computer Use, MolmoWeb-8B, GPT-5.4); several rows are marked Verified=False (self-reported). A separate human-eval table (13 rows, leader Hark 97.7) is on the same Space. Frontier models of Oct 2026 are not listed.
Qwen3.7-Plus2.8% on OSWorld 2.0 (end-to-end models)$–
Claude Fable 5.1No public result yet–
Gemini 3.8 FlashNo public result yet–
GPT-6 AstraNo public result yet–
n1.5No public result yet
MiniMax-M3
EvoCUA-32B-20260105
Fara1.5 (27B)
GUI-Owl-1.5 (32B)
Holo4 (35B-A3B)
MolmoWeb-8B
OpenCUA-72B
UI-TARS-1.5-7B
Sai (Simular)
ChatGPT
Claude
Comet
Dia
Fellou
Gemini
Genspark
Manus
Opera Neon
Nanobrowser
MagenticLite (formerly Magentic-UI)
UiPath
Stagehand
Amazon Nova Act
Microsoft Copilot Studio
OpenClaw
Hermes Agent
Browser Use
UI-TARS Desktop
Cua
Skyvern
Midscene.js
Agent S
UFO³