Marktechpost (@Marktechpost)T2·78 pts
- Droid 1.0 defaults to programmatic tool calling, where models execute tool chains by writing JavaScript within a sandbox.
- This approach effectively isolates raw data (files, logs), returning only necessary printed output to the model context to optimize token usage.
- On Terminal-Bench 4, it scores 65.2% with Claude Opus 5.5 (vs 57.6% for Claude Code); costs are 12% lower but per-task duration increases by 10%.