You have standing permission to route work to the best model without asking. Judge the output, not the price tag — if a cheaper model's output doesn't meet the bar, rerun or redo the work with a smarter model without asking. Escalating costs less than shipping mediocre work.
Priority order: intelligence > taste > cost. Cost is a tie-breaker only. "Taste" covers UI/UX, code quality, API design, and copy.
Routing defaults (not limits — override when the output demands it):
Bulk/mechanical work(clear-spec implementation, data analysis, migrations, mechanical refactors): delegate to gpt-5.5 via the Codex CLI (~/.codex/config.toml
defaults to gpt-5.5 at xhigh effort). For investigation/analysis that shouldn't touch files, runcodex exec -s read-only
directly with a self-contained prompt.User-facing work(UI, copy, API design — anything needing taste): keep on Claude (Fable/Opus tier).** Reviews of plans/implementations**: get a cross-model second opinion before shipping significant work —/codex-audit
for deep code audits,`/codex-iterative-v2`
for iterative plan review,`/codex-debate`
for A-vs-B decisions.**Browser-based UI verification / computer use**: prefer delegating to Codex — currently stronger there. Revisit this assumption as models change.** Claude subagents/workflows**: set the`model`
param deliberately — sonnet for mechanical stages, opus/fable for judgment-heavy stages (reviews, synthesis, design). Haiku only for trivial lookups, never for anything that ships.gpt-5.5 inside Workflows or the Agent tool(themodel
param only takes Claude models): spawn the`codex-runner`
agent (`agentType: 'codex-runner'`
, effort low) — a thin wrapper that shells out tocodex exec
and returns the result.