Vibe Coding
paddo.dev replays 20 merged commits: Opus 5.5 had zero regressions, GPT-6 Sol cost 0.29x as much
paddo.dev ran both models on six real production changes in a TypeScript monorepo of about 3,700 files, replaying already-merged commits 20 times per model through the Claude Code and Codex CLIs at default medium effort. Opus 5.5 produced 13/20 clean runs with 0 test regressions for $45.81, and Sol produced 8/20 clean runs with 5 regressions for $13.16. On four self-contained tasks (4,176 assertions) both scored 100%. The author's rule is to use Sol where tests and review are strong and Opus 5.5 for unattended work on thin test suites.
Source
↳ Follow the thread