Hacker News AIT2·78 pts
- Switching models warrants re-evaluating the harness; each model has a different best-value harness
- Cost-per-task rankings shift by a median of 11 of 22 places, with Codex CLI dropping from 1st to 16th
- 64% of inference spend went to failed runs; step count, not tool count or cache hit rate, predicts cost per pass