Sai computer-use agent scores 73% on OSWorld 2.0, beating Opus 5 at lower cost
Med1_Ai · x · 2026-10-21
The Sai team demoed its computer-use agent operating a Windows PC to scan the Meta Ad Library for sneaker-category ads, autonomously searching Nike, Adidas and New Balance.
- Of 14 ads returned: 8 were retailer assortment pushes with no discounts, 5 mentioned free shipping, and most New Balance results were one Amazon template repeated; the agent self-caught 2 wrong entries before sending
- The run used 233 credits and 15 minutes of computer time; setup is one line in Claude Code
Benchmarks: on OSWorld 2.0, Sai scores 73.0% at $15.70/task, ahead of Opus 5 (70.6% at $23.70) and GPT-5.6 Sol (62.6% at $26.62).
More from Apps
- 1Password launches tool letting AI agents log in for you without seeing your password — AndrewSchmidtFC · 2026-10-23
- Lamin goes GA: open-source data platform packs a dozen bio-AI tools into one Python package — ycombinator · 2026-10-23
- Ethan Mollick shares a useful Codex prompt for auditing your computer security — emollick · 2026-10-23
- Escalate details plan for binder-based combinatorial proteomics to feed AI biology models — PatrickKidger · 2026-10-23
- Boardy AI chief of staff refuses to fake data, inspires a smarter market-ranking method — chickefitz · 2026-10-23
- LokalBot ships Linux/Windows alpha: open-source local transcription, dictation and autocomplete — stevyhacker · 2026-10-23