Writing
Why I moved my Codex default to GPT-6.1 Sol High
GPT-6.1 Sol at high led the DeepSWE scores I checked, but the measured gains vary by benchmark. Here is why I changed my default and how I adjusted the delegation rules around it.
Sonnet 5.5: read the curve, not the headline
The headline said Sonnet 5.5 beats Opus 5.5. At matched spend it doesn't, so my main session stayed put. The same curves showed why my grunt-work agent no longer belongs at medium.
Three models, three jobs: my GPT-6 Codex setup
I moved the Codex team to GPT-6: Sol coordinates, Astra takes the rare hard calls, and Luna handles bounded execution. The roles are policy, not an automatic model optimizer.
Three charts, not one: moving my default to Opus 5.5
Opus 5.5 finally made the "switch the main session off Fable" lever worth pulling. The first benchmark chart said medium effort; three charts said high.
The dial that moved the wrong thing
For six weeks my session effort dial was also, silently, my deep-reasoning subagent's dial. The argument for turning it down assumed it wasn't. One line of frontmatter, and a free way to check which cap your subagents bill.
Fable 5.1: three places a model ID lives
I changed one line to move to Fable 5.1 and thought I was done. A shell wrapper I'd written myself was quietly overriding it — and the same class of silent failure was waiting in my app code.
Medium by default: when the token math stops adding up
An App Store review pass and one policy site burned 69% of a weekly Fable cap in ten hours. I don't know why yet, so the effort default came down to medium as a holding action.
Effort economics on Fable 5: coming down from xhigh
A day after un-pinning Ultracode, the effort default came down another notch — from xhigh to high. Why lower effort on Fable 5 isn't the downgrade it sounds like.
How I Set Up Codex as a Small Model Team
The small Codex setup I use: Sol orchestrates, Terra handles deep reasoning, and Luna handles focused implementation work.
Un-pinning Ultracode: a CLAUDE.md update
Why I stopped running Ultracode as a standing default, the per-stage effort rules that replaced it, and the two-command fallback when the Fable cap runs out.
My actual CLAUDE.md, annotated
My real global CLAUDE.md, translated line for line, with notes on why each rule is there.