Mage-VL-Cowork Card - OSWorld-Verified (361 tasks, 50 steps) - Daily: leaderboard
OSWorld-Verified desktop computer-use tasks in real applications, restricted to the 361 tasks left after excluding the Google Drive tasks, with a 50-step budget, as run by Microsoft for the Mage-VL-Cowork card (2026-10-30): Mage-VL-Cowork (4B) in the MAGEHybridAgent harness (GUI actions plus a bash channel, the 3 most recent screenshots, 1920x1080) as the mean of two full runs, and GPT-5.4-mini at xhigh effort in one run with its harness not stated. Other rows of the card's table are reprints or other step budgets and are not on this board. This board is the Daily domain (daily-use application tasks). Task success (%), higher is better.
Metric: Success rate (%). Source: huggingface.co. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini (xHigh) | 76.6 | #228 (GPT-5.4 Mini) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mage-vl-cowork-card-osworld-verified-361-tasks-50-steps-daily · How It Works · Data refreshed daily, snapshot 2026-11-05.