LLM Stats leaderboard
44 scores · 44 models · first recorded Oct 2, 2026
higher is better · scale 0–1 · record OSWorld
Best reported score
86.1%Qwen3.8 Max · LLM Stats leaderboard
47 models tested · 2 sources · 8 source documents · latest report Aug 14, 2026
[NeurIPS 2024] OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
Scores stay separate per source — never merged into one cross-source ranking.
44 scores · 44 models · first recorded Oct 2, 2026
higher is better · scale 0–1 · record OSWorld
| Model | Value (%) | Reported | Score type | Recorded |
|---|---|---|---|---|
| Qwen3.8-27B | 84.3% | Aug 14, 2026 | unspecified | Oct 2, 2026 |
| Muse Glimmer-30B | 65.9% | Aug 10, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.8 Max | 86.1% | Aug 2, 2026 | unspecified | Oct 2, 2026 |
| Gemini 3.6 Flash | 83% | Jul 21, 2026 | unspecified | Oct 2, 2026 |
| Gemini 3.5 Flash-Lite | 74% | Jul 21, 2026 | unspecified | Oct 2, 2026 |
| Muse Spark 1.1 | 80.8% | Jul 9, 2026 | unspecified | Oct 2, 2026 |
| Claude Sonnet 5 | 81.2% | Jun 30, 2026 | unspecified | Oct 2, 2026 |
| Seed 2.1 Pro | 78.8% | Jun 24, 2026 | unspecified | Oct 2, 2026 |
| Seed 2.1 Turbo | 76.4% | Jun 24, 2026 | unspecified | Oct 2, 2026 |
| Claude Fable 5 | 85% | Jun 9, 2026 | unspecified | Oct 2, 2026 |
| MiniMax M3 | 70.1% | Jun 1, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.7-Plus | 73.3% | May 31, 2026 | unspecified | Oct 2, 2026 |
| Claude Opus 4.8 | 83.4% | May 28, 2026 | unspecified | Oct 2, 2026 |
| Gemini 3.5 Flash | 78.4% | May 19, 2026 | unspecified | Oct 2, 2026 |
| GPT-5.5 | 78.7% | Apr 23, 2026 | unspecified | Oct 2, 2026 |
| Kimi K2.6 | 73.1% | Apr 20, 2026 | unspecified | Oct 2, 2026 |
| Claude Opus 4.7 | 78% | Apr 16, 2026 | unspecified | Oct 2, 2026 |
| Claude Mythos Preview | 79.6% | Apr 7, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.6 Plus | 62.5% | Apr 2, 2026 | unspecified | Oct 2, 2026 |
| GLM-5V-Turbo | 62.3% | Apr 2, 2026 | unspecified | Oct 2, 2026 |
| GPT-5.4 mini | 72.1% | Mar 17, 2026 | unspecified | Oct 2, 2026 |
| GPT-5.4 nano | 39% | Mar 17, 2026 | unspecified | Oct 2, 2026 |
| GPT-5.4 | 75% | Mar 5, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.5-122B-A10B | 58.0% | Feb 24, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.5-27B | 56.2% | Feb 24, 2026 | unspecified | Oct 2, 2026 |
| Qwen3.5-35B-A3B | 54.5% | Feb 24, 2026 | unspecified | Oct 2, 2026 |
| Claude Sonnet 4.6 | 72.5% | Feb 17, 2026 | unspecified | Oct 2, 2026 |
| GPT-5.3 Codex | 64.7% | Feb 5, 2026 | unspecified | Oct 2, 2026 |
| Claude Opus 4.6 | 72.7% | Feb 5, 2026 | unspecified | Oct 2, 2026 |
| Claude Opus 4.5 | 66.3% | Nov 24, 2025 | unspecified | Oct 2, 2026 |
| Claude Haiku 4.5 | 50.7% | Oct 15, 2025 | unspecified | Oct 2, 2026 |
| Claude Sonnet 4.5 | 61.4% | Sep 29, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 235B A22B Instruct | 66.7% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 32B Thinking | 41% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 235B A22B Thinking | 38.1% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 8B Thinking | 33.9% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 8B Instruct | 33.9% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 32B Instruct | 32.6% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 4B Thinking | 31.4% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 30B A3B Thinking | 30.6% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 30B A3B Instruct | 30.3% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen3 VL 4B Instruct | 26.2% | Sep 22, 2025 | unspecified | Oct 2, 2026 |
| Qwen2.5 VL 32B Instruct | 5.92% | Feb 28, 2025 | unspecified | Oct 2, 2026 |
| Qwen2.5 VL 72B Instruct | 8.83% | Jan 26, 2025 | unspecified | Oct 2, 2026 |
4 scores · 3 models · first recorded Oct 2, 2026
higher is better · success · scale 0–100 · record OSWorld
| Model | Value (%) | Subset / variant | Test conditions | Reported by | Reported | Document |
|---|---|---|---|---|---|---|
| Kimi K3 | 84.8% | osworld_verified | reasoning=max | Moonshot AI | Jun 13, 2026 | moonshot_kimi_k3_model_card |
| Claude Opus 5 | 70.6% | osworld_2.0 | max effort, first-attempt success rate, averaged over 5 runs, 1080p, max 500 action steps | Anthropic | Jul 24, 2026 | anthropic_claude_opus_5_system_card |
| Qwen3.5-397B-A17B | 62.2% | osworld_verified | thinking (temp 0.6, top_p 0.95, top_k 20) | Qwen | Feb 16, 2026 | qwen3_5_model_card |
| Kimi K3 | 58.3% | osworld_2.0 | reasoning=max | Moonshot AI | Jun 13, 2026 | moonshot_kimi_k3_model_card |
| Model | Value (%) | Reported | Score type | Recorded |
|---|---|---|---|---|
| Claude Opus 5 | 70.6% | Jul 24, 2026 | self-reported | Oct 2, 2026 |
| Kimi K3 | 58.3% | Jun 13, 2026 | self-reported | Oct 2, 2026 |
| Kimi K3 | 84.8% | Jun 13, 2026 | self-reported | Oct 2, 2026 |
| Qwen3.5-397B-A17B | 62.2% | Feb 16, 2026 | self-reported | Oct 2, 2026 |
every document that defines or reports scores
| other | – | OSWorld– |
| release post | Anthropic | Claude Fable 5 and Mythos 5Jun 9, 2026 |
| system card | Anthropic | Claude Opus 5Jul 24, 2026 |
| other | Moonshot AI | Kimi K3Jun 13, 2026 |
| other | Qwen | Qwen3.5-397B-A17BFeb 16, 2026 |
| other | Qwen | Qwen3.8-27BAug 5, 2026 |
| other | – | ubuntu_osworld– |
| other | – | OSWorld-Verified– |
Records are located via the tracking corpus (ktwu01/benchmark-radar live dataset). Every row cites its source document.