Three frontier models in a row see users rolling back to older versions

alexisgallagher · x · 2026-09-12

@kunchenguid tracks user sentiment and reports three consecutive frontier model launches where many users rolled back: Opus 5 (widely negative — poor communication, jumping to conclusions, repeat mistakes, even a dedicated complaint site; users returned to 4.6/4.8) and GPT 6 Astra (users reverting to 5.6 due to high cost as a daily driver and spiky performance).
@alexisgallagher argues models are genuinely improving, but slower than user demands are growing: as models generate more code, users need them to handle global architectural coherence in larger codebases and parallel tasks — hence the focus on chief-of-staff patterns and embedding architectural invariants into repos.
Some speculate excessive RL may explain the perceived slowdown in quality gains.

More from Models