Turning Opus from a Grunt into an Advisor: This Claude Code Hybrid Setup Saves a Chunk
The advisor feature in Anthropic's docs was configured by xsmryze so that Sonnet does the main writing, Haiku explores, and Opus only gives judgments at key points. One command, with prerequisites for saving money.
Anthropic's docs have an experimental feature called advisor that lets the main model consult a stronger model at key points. In a tweet, xsmryze configured it into a more economical division of labor: Sonnet does the main writing, Haiku explores, and Opus only serves as advisor. The launch command is just one line:
claude --model sonnet --advisor opus
The division of labor is: Sonnet 5.5 runs the main session and edits code; Haiku 5.5 searches files, greps call sites, and pulls API docs in parallel; Opus 5.5 stays online waiting for the main model to ask. xsmryze summarizes Opus's intervention points into three: before a plan is locked, when the same error appears twice, and before declaring a task complete. The docs are more conservative: Claude decides for itself when to call the advisor, usually at these points, with no hard-coded rules.
The advisor receives the full conversation, including tool calls and results, then returns advice. It is billed at the advisor model's rate, and every call rereads the full record. Turning the advisor on or off does not invalidate the main model's prompt cache, but the advisor's own reads are not cached. It suits long tasks: most turns are routine, and plan quality determines the result. For short tasks, or work that needs the strongest model every turn, just switching the main model is simpler.
There is an easily overlooked prerequisite here. Don't treat the advisor as an automatic money-saving button. The docs state it plainly: the advisor only takes effect when it is asked. In a long post, NO1ennn cites Anthropic's recommendation: for most coding and knowledge work, start with Opus 5.5, and move to Fable when the task is large, long, and open-ended. He also lists the cost: Fable costs twice as much per token as Opus, four times as much as Sonnet, and 80 times as much as Haiku. Expensive models make judgments, cheap models execute; the logic holds. But the pairing method should be validated with /usage and results, not just by looking at unit price.
In a reply, Synra posted a game made with a single HTML file and said the whole thing was done before 5.2 was released. Such cases can show results, but don't treat a demo as proof of production readiness.
Model pairing also has hard limits. The advisor cannot be less capable than the main model; Haiku 4.5 can call an advisor but cannot serve as one; Fable 5.2 requires Claude Code v2.1.321 or newer. On some plans, using Fable as an advisor goes through usage credits and requires one-time authorization. When the main model is Fable, the advisor can only be Fable 5.2, Fable 5.1, or Fable 5. These are clearly laid out in the docs table.
xsmryze also provided a prompt to have Claude Code rebuild the configuration around this tree. You can first use diff to review the changes before deciding whether to execute:
Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go.
The tweet also mentions that JEV handles micro-branches such as tool choice, path selection, and retries, compressing them to milliseconds. This part is his engineering solution; the advisor docs do not list it as mandatory. For ordinary users, getting Sonnet plus an Opus advisor working is enough.
Docs link: http://code.claude.com/docs/en/advisor
Long-post link: https://x.com/i/article/2120515552300941313
Original tweet: https://x.com/xsmryze/status/2123192904264913342
发布时间: 2026-11-19 08:28