277 runs of Claude Code, Codex, and pi on bug fixes from my own repos: an 80x cost spread, speed sorted by harness, and a sandbox setting that hid failures behind green scores.

4 points•ariwilson•6 days ago•1 comment•

1 comment

jboggan6 days ago
I'm more excited about the cost and speed of "good enough" for models like Luna than I am for the frontier models. A great write-up, thanks.

Read the full thread on Hacker News →

Related stories