Benchmark leaderboard - 25 entries
OSWorld leaderboard
OSWorld is a benchmark of 369 computer tasks involving real web and desktop apps in open domains, OS file I/O, and workflows spanning multiple applications, each with an initial state setup and a custom execution-based evaluation script, run in a real computer environment. As of October 17, 2026, claude-fable-5[1m] (Anthropic) ranks first on the OSWorld (OSWorld-Verified) board with 86.0%.
Last verified from osworld-v1.xlang.ai. Read the caveats
OSWorld (OSWorld-Verified)
25 entries, in the board's order
| 1 | claude-fable-5[1m] (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 307.73 / 358; rank shown as 1 with an asterisk on the leaderboard Official leaderboard | 86.0% |
| 2 | claude-opus-5[1m] (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 300.2 / 360 Official leaderboard | 83.4% |
| 3 | Muse Spark 1.1 (Meta Superintelligence Labs) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 291.22 / 361 Official leaderboard | 80.7% |
| 4 | Holo3-35B-A3B (H Company) OSWorld-Verified Results, Type: Specialized model, Max Steps: 100, Runs: 2 (Run 1: 296.41/359, Run 2: 280.55/359) Official leaderboard | 80.4±2.2% |
| 5 | MiniMax M3 (MiniMax) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 269.18 / 358 Official leaderboard | 75.2% |
| 6 | Qwen 3.7 Plus (Qwen Team, Alibaba Group) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1 Official leaderboard | 73.3% |
| 7 | Kimi K2.6 (Moonshot AI) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 261.58/358 Official leaderboard | 73.1% |
| 8 | claude-sonnet-4-6 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 256.71/356 Official leaderboard | 72.1% |
| 9 | Kimi K2.5 (Moonshot AI) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 228.65/361 Official leaderboard | 63.3% |
| 10 | claude-sonnet-4-5-20250929 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 226.35/360 Official leaderboard | 62.9% |
| 11 | Seed-1.8 (ByteDance Seed) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 222.12/359 Official leaderboard | 61.9% |
| 12 | claude-sonnet-4-5-20250929 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 50, Runs: 1, Success/Total 209.68/361 Official leaderboard | 58.1% |
| 13 | EvoCUA-20260105 (Meituan LongCat Team) OSWorld-Verified Results, Type: General model, Max Steps: 50, Runs: 1, Success/Total 204.23/360 Official leaderboard | 56.7% |
| 14 | GUI-Owl-1.5 32B (Tongyi Lab, Alibaba Group) OSWorld-Verified Results, Type: Specialized model, Max Steps: 50, Runs: 1, Success/Total 197.37/356 Official leaderboard | 55.4% |
| 15 | DeepMiner-Mano-72B (Mininglamp Technology) OSWorld-Verified Results, Type: Specialized model, Max Steps: 100, Runs: 1, Success/Total 194.62/361 Official leaderboard | 53.9% |
| 16 | UI-TARS-2-2509 (ByteDance Seed) OSWorld-Verified Results, Type: Specialized model, Max Steps: 100, Runs: 1, Success/Total 191.69/361 Official leaderboard | 53.1% |
| 17 | EvoCUA (Meituan LongCat Team) OSWorld-Verified Results, Type: General model, Max Steps: 50, Runs: 1, Success/Total 181.06/360 Official leaderboard | 50.3% |
| 18 | EvoCUA-8B-20260105 (Meituan LongCat Team) OSWorld-Verified Results, Type: General model, Max Steps: 50, Runs: 1, Success/Total 166.28/361 Official leaderboard | 46.1% |
| 19 | opencua-72b-preview (The University of Hong Kong & Moonshot AI) OSWorld-Verified Results, Type: Specialized model, Max Steps: 100, Runs: 3 (Run 1: 165.89/360, Run 2: 157.40/359, Run 3: 161.73/359) Official leaderboard | 45.0±0.9% |
| 20 | opencua-72b-preview (The University of Hong Kong & Moonshot AI) OSWorld-Verified Results, Type: Specialized model, Max Steps: 50, Runs: 1, Success/Total 160.7/358 Official leaderboard | 44.9% |
| 21 | claude-4-sonnet-20250514 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 50, Runs: 1, Success/Total 158.42/361 Official leaderboard | 43.9% |
| 22 | claude-sonnet-4-5-20250929 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 15, Runs: 1, Success/Total 154.80/361 Official leaderboard | 42.9% |
| 23 | UI-TARS-250705 (ByteDance Seed) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 151.06/361 Official leaderboard | 41.8% |
| 24 | qwen3-vl-flash-2025-10-25 (Qwen Team, Alibaba Group) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 150.07/361 Official leaderboard | 41.6% |
| 25 | claude-4-sonnet-20250514 (Anthropic) OSWorld-Verified Results, Type: General model, Max Steps: 100, Runs: 1, Success/Total 149.41/361 Official leaderboard | 41.4% |
About this board
- Entries are the OSWorld-Verified Results tab (official, evaluated by the OSWorld team under unified settings), sorted by score; the leaderboard lists the same model multiple times at different Max Steps (15/50/100), so the top 25 rows include duplicate models at different step budgets (claude-sonnet-4-5-20250929 x3, opencua-72b-preview x2, claude-4-sonnet-20250514 x2).
- Self-reported results tab (non-verified) was not captured.
FAQ
What does OSWorld measure?
OSWorld is a benchmark of 369 computer tasks involving real web and desktop apps in open domains, OS file I/O, and workflows spanning multiple applications, each with an initial state setup and a custom execution-based evaluation script, run in a real computer environment.
Which model leads OSWorld?
As of October 17, 2026, claude-fable-5[1m] (Anthropic) ranks first on the OSWorld (OSWorld-Verified) board with 86.0%. That is the board's own order, not a Developers Digest ranking.
Where do these OSWorld scores come from?
All 25 entries come from the official leaderboard at osworld-v1.xlang.ai. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- Muse Spark Open Weights: 816B Model, Benchmarks, and How to Run ItMeta released Muse Spark Open weights on October 15, 2026: 816B total, 42B active, 1M context. Run recipes, hardware floor, license catch, and where it trails.
- Cheapest Subagent Model: Haiku 5.5 or GPT-6 Luna? Routing, Compaction and Cost per TaskWhich cheap model belongs under your orchestrator? Haiku 5.5 vs GPT-6 Luna as a subagent or worker: routing by prompt length, compaction, tool-call gotchas and cost per agent task.
- Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5: Which Claude Model to UseSonnet 5.5 is the default, Opus 5.5 is for open-ended long-horizon work, and Haiku 5.5 is for bounded steps and subagents. Prices, benchmarks, cost math.
- Claude Haiku 5.5: Pricing, Migration Changes, and When to Use ItClaude Haiku 5.5 pricing, the cliff above 100K-token prompts, the tokenizer catch, the migration steps from Haiku 4.5, and when to use it instead of Sonnet.
- OpenAI DevDay 2026 Recap: What Shipped and What to TestEverything OpenAI announced at DevDay 2026: GPT-6.1 Sol at $2/$10, Ultrafast, Decisions API, Agents API, cloud Codex, and what to test first.
- GPT-6.1 Sol Release Guide: Near-Astra Agentic Work at $2/$10 and $0.10 Cache ReadsGPT-6.1 Sol release guide: near-Astra scores at a fifth of Astra's price, $0.10 cache reads, verified pricing, the Astra 6.1 safety hold, and how to run it.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.