Computer-use agent API claims 90% token savings, tops OSWorld 2.0 at 73% for $15.70 per task
AdFluid9823 · reddit · 2026-10-04
A newly opened computer-use API called Sai targets the pain points of screenshot-based agents: per-step reasoning is slow, token-hungry, and brittle to UI changes.
- Instead of re-planning every step, Sai compiles plans into self-healing code that survives UI changes; repeated tasks rerun that code instead of reasoning through each screen, saving up to 90% of tokens.
- On OSWorld 2.0 (108 hard real desktop tasks), Sai scored 73% at $15.70 per task, versus Claude Opus 5 at 70.6% / $23.70 and GPT-5.6 Sol at 62.6% / $26.62.
- Use cases: GUI work with no API—legacy desktop apps, Excel, internal tools, form-heavy portals, data entry. Free plan includes a cloud computer.
- The research behind it is open source: Agent S (ICLR 2025, 12k GitHub stars), the first agent to pass human-level performance on OSWorld.
The team is soliciting feedback and use cases from developers building agents.
More from coding & agent
- AI agent negotiates car lease and mortgage, saving 15-20% by voice — armand_ruiz · 2026-10-04
- Teknium says setting up Hermes with Discord is now less confusing and painful — Teknium · 2026-10-04
- Asking ChatGPT to remove a feature somehow adds 231 lines of code — rickasaurus · 2026-10-04
- Jensen Huang explains agent harness: the exoskeleton that makes LLMs useful — rohanpaul_ai · 2026-10-04
- Windows Defender flags 12 filenames in a test DLL; a JSON repack clears scans — bytebot · 2026-10-04
- Vitalik's privacy-preserving personal AI: local Qwen orchestrator + zkAPI + Tor, no data leaks — kenziyuliu · 2026-10-04