TL;DR
The short version
Claire Vo left Claude because it was annoying, not dumb. Opus 5.5 clears that bar — cleaner talk, stronger front-end/SVGs, solid long agentic runs — while Codex keeps daily harness muscle memory.
On How I AI (22 Sep 2026), list pricing is $4/$20 per MTok (20% under Opus 5’s list, not a flat “40% cheaper” sticker story).
Distilled from Claire Vo’s How I AI review Claude is BACK with Opus 5.5.
The bar was “is it annoying?”
Vo had abandoned Claude’s harness because models rambled. Opus 5.5 passes her personality gate. When she thought she was annoyed at 5.5, she had accidentally been on Opus 5.
I stopped using Claude not because of its intelligence or its models. I stopped using Claude cuz it was annoying.
Claire Vo
Price, quiet latency, and real work
API list is $4/$20 per million tokens; fast mode $8/$40. That is a 20% list cut versus Opus 5 — Anthropic’s separate ~40% claim is workload/token efficiency. Meter your own runs.
Quieter replies cut annoyance and create “are you still working?” dead air on long turns. Her inbox/backend/research/computer-use suite completed; she now uses Opus as an adversarial PR reviewer beside Codex.
Harness still beats model for default
First Opus with Fable-class cyber/bio rails will refuse a yolo prod push. Safety posture leaks into everyday personality.
Operating split: Opus for PR review, architecture, front-end/SVG; Codex for computer use, video, and dock habit. “Claude is back” ≠ “Claude is default.”
