Paid AI tools silently regress between versions — one user's golden test case defense

Ok-Independent3290 · reddit · 2026-09-18

A heavy-paying user of ChatGPT Pro and Claude Max reports that an AI document workflow used for months started dropping formatting and ignoring instructions after an update — with no changelog entry or announcement.

The real cost, the author argues, isn't rate limits or known model swaps (those can be planned around) but undocumented backsliding: silent regressions force re-testing prompts you already validated.

Mitigation: assume any agent you depend on will subtly degrade at some point, and keep a few golden test cases to re-run after every update.

Original post →

More from coding & agent

coding & agent channel →