Claude Code's undocumented --advisor flag pairs Sonnet with Opus
A hidden --advisor flag in Claude Code's docs lets Sonnet run the whole session while Opus is consulted only at key decision points, avoiding Opus rates on every turn.
WHOEVER SHIPPED THIS FLAG AND LEFT IT OUT OF --help HAS BIGGER BALLS THAN SENSE claude --advisor opus 1 flag. buried in Anthropic's own Claude Code docs. NOT in claude --help. run ONE model and you pay its rates on EVERY SINGLE turn. the fix is sitting right in the docs: the pairing. The idea: Sonnet 5.5 runs the whole session. Opus 5.5 sits on call and only gets pulled in at decision points, not on every turn. launch it like this: claude --model sonnet --advisor opus here's the split: → Sonnet 5.5 drives the session and writes the code → Haiku 5.5 subagents explore the repo in parallel, read-only → Opus 5.5 reads the FULL transcript, every tool call, every result, the moment it's consulted the 3 moments Opus is built to catch: → before the plan locks → when the same error shows up twice → before Claude calls it done Sonnet builds. Haiku explores. Opus reviews. 3 things nobody tells you: → flipping advisor on or off mid-session does NOT invalidate your prompt cache → swap Opus for Fable 5.1 as the advisor (needs Claude Code v2.1.257+) → Haiku 4.5 can CALL an advisor but can NEVER be one that's the difference between paying Opus on every routine turn and paying Opus only on the calls that decide the outcome. paste this into Claude Code ↓ "You're running on Sonnet 5.5 with Opus 5.5 as advisor. Spin up 3 read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference). Consult the advisor before you lock a plan, when an error repeats, and before calling a task done. Show me every change as a diff first." ↳ http://code.claude.com/docs/en/advisor save it. you'll use it every single day.
This post quotes another post: x.com/plusxbt/status/2110106944258425061