CHOOSEAgents & automation
Computer-Use Agent Cost Optimization
creedants@creedants
This project optimizes the cost of computer-use agents by shifting from complex planning to closed-set choice, significantly reducing per-decision costs.
TakeWe tried the same shape on real pages: pick from on-screen options, no screenshot model.
Most of the bill in computer-use agents is not the click.
GTA1 put average task cost at $2.43, and ~87% of that was planning / reflection, not actuation.
The fix pattern is closed-set Choice instead of a frontier plan every step. TypeSafe’s public computer-use build clocks Jev at ~$0.0002 per decision vs ~$0.032 for Opus on a bare screenshot (~160× cheaper per pick). On a 12-step task that’s about $0.003 vs $0.40–$0.90.
We tried the same shape on real pages: pick from on-screen options, no screenshot model. Confidence landed 0.88–0.96 on public sites and 0.99 on a controlled dry run. When the next move is a short list, you’re mostly paying for a classifier. Planning tax shows up when the list is open.
Screenshot agents pay for plans. Classifier agents pay for picks.
Related builds
