Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresFrontierCUA v1, Collinear AI's private computer-use benchmark of 40 held-out connected workflows across running clones of enterprise applications such as HubSpot, Zendesk, Slack and Google Sheets; agents work only from the screen, a screenshot and one mouse or keyboard action per step, and an AI judge scores each workflow's weighted rubric of yes-or-no checks from the recorded final application state. Each model has one graded attempt per workflow; an attempt passes at a grade of at least 0.95, and pass@1 is the share of workflows passed. A new task cohort, kept separate from the 15-workflow v0 snapshot.
267th strongest of 818 fitted benchmarks
Signal-to-noise +8.69
50th hardest of 818 fitted benchmarks
Capability needed ≳ 179.5
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high | 42.5 | |
| 2 | GPT-6.2 Astra | high · Codex | 32.5 | |
| 3 | Claude Fable 5.2 | high | 25.0 | |
| 4 | GPT-6.1 Sol | high · Codex | 25.0 | |
| 5 | GPT-6.2 Sol | high · Codex | 25.0 | |
| 6 | GPT-6 Astra | high · Codex | 20.0 | |
| 7 | Claude Opus 5 | high | 17.5 | |
| 8 | MiniMax M3.1 Flash | max | 17.5 | |
| 9 | Qwen4-Max | high · Qwen-Agent computer-use harness (OSWorld-V2 QwenAgent port) via OpenRouter | 17.5 | |
| 10 | Claude Fable 5.1 | high | 15.0 | |
| 11 | Claude Fable 5 | high | 15.0 | |
| 12 | Claude Opus 5.5 | high | 12.5 | |
| 13 | Sonnet 5 | high | 12.5 | |
| 14 | Kimi K3 | high | 12.5 | |
| 15 | MiMo-V2.6-Flash | high | 12.5 | |
| 16 | Muse Wonder | high | 12.5 | |
| 17 | GPT-5.6 Sol | high · Codex | 7.5 | |
| 18 | GPT-6.1 Terra | high · Codex | 7.5 | |
| 19 | GPT-6 Sol | high · Codex | 7.5 | |
| 20 | Grok 4.7 | high | 7.5 | |
| 21 | Muse Ember | high | 7.5 | |
| 22 | MiMo-V2.6-Pro | high | 5.0 | |
| 23 | GLM-5.3 Flash | high · GLM-5.3-FlashX inference-speed tier via Alibaba Cloud | 2.5 | |
| 24 | Kimi K3.1 Pro | high | 2.5 | |
| 25 | Muse Spark 1.3 | high | 0.0 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).