We counted it. Real operator work, two machines.
Opus 5: 14.0 self-corrections per 1k
Fable 5.1: 4.2
3.3× more walk-backs. Not more wrong — more unsupported first claims.
benchagi.com/s/gTF-NlOZ5pxYe…
cc @bcherny @alexalbert__ @AmandaAskell
Sep 20, 2026 · 5:24 AM UTC
1
1,565
