The Claude Code advisor lets one model do the work while a stronger model is consulted only at the moments that decide the outcome. With Opus 5.5 as the main model and Fable 5.2 as the advisor, you get Fable's judgement on the plan, on recurring errors and on the final check, without running Fable on every turn.
What the Claude Code advisor does {#what}
The advisor is a server tool that runs on Anthropic's side. At a key moment the main model calls it; the advisor model receives the full conversation, every tool call and result included, and returns guidance. The main model applies that guidance and keeps working. The advisor never edits files itself, and when the main model's own evidence contradicts the advice (a recommended step fails, a file says otherwise), it surfaces the conflict instead of following blindly (Claude Code docs).
How to set Fable 5.2 as the advisor {#setup}
There are three ways, from the docs:
# 1. Inside a session: set it and save it as your default /advisor fable # 2. In your settings file (~/.claude/settings.json): a persistent default { "advisorModel": "fable" } # 3. For one session only, without changing the saved setting claude --advisor fable fable, opus and sonnet are aliases for Claude Code's current default in each family; a full model id such as claude-opus-5-5 also works. To check what is active, run /advisor with no argument. On our Mac (Claude Code 2.1.322) it printed this before and after adding the setting:
$ claude -p "/advisor" Advisor: off Usage: /advisor <fable|opus|sonnet|off> # after adding "advisorModel": "fable" to ~/.claude/settings.json $ claude -p "/advisor" Advisor: Fable 5.2 Usage: /advisor <fable|opus|sonnet|off> On some plans, Fable as the advisor first needs a one-time consent: run /model fable and continue on Fable once, then pick Fable as the advisor. Until then Claude Code points you to /model fable and does not save Fable as the advisor.
Which advisor works with which model {#pairing}
The rule: the advisor must rank at or above the session's main model. Claude Code checks the pairing before each request, and the API checks it again.
Accepted pairings, from the Claude Code docs. Captured November 10, 2026. Source | Main model | Accepted advisor | Note |
|---|---|---|
| Opus 5 or Opus 5.5 | Fable, Opus 5 or later | Opus does the work, Fable advises |
| Sonnet 5.5 | Fable, Opus 5 or later, Sonnet 5.5 | Fable advice without running Fable throughout |
| Fable 5.2 | Fable 5.2 only | Claude Code does not attach an Opus or Sonnet advisor to Fable |
| Haiku 4.5 | Fable, Opus, Sonnet, Haiku 5.5 | Haiku 4.5 can call an advisor but cannot be one |
Subagents inherit the advisor and run the same check against their own model. If you pick an advisor below your main model, Claude Code saves it but does not attach it.
When Claude asks the advisor {#when}
Claude decides, not a fixed rule. In practice it tends to consult:
- before committing to an approach,
- when the same error keeps coming back,
- before declaring the task done.
You can ask for one directly ("consult the advisor before you continue"). There is no setting that caps or forces calls; say in your instructions if you want more or fewer. During a call the transcript shows an Advising line; press Ctrl+O to read the guidance. Turning the advisor on or off mid-session does not break the main model's prompt cache.
Advisor vs opusplan vs subagents {#compare}
When the stronger model runs in each approach, from the Claude Code docs. Captured November 10, 2026. Source Use the advisor for long tasks where most turns are routine but the plan decides the outcome: large refactors, a bug that keeps returning, work you want checked before it is called done. If you already run Codex CLI next to Claude Code for a second opinion, the advisor gives you one inside the Claude session. For short tasks, or work where every turn needs the strongest model, switch the main model with /model instead. To hand a whole subtask to another model, use a subagent with its own model.
Requirements and limits {#limits}
- Experimental, and only on the Anthropic API or Claude Platform on AWS (there it needs Claude Code 2.1.300 or later): not on Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry (requirements).
- Fable 5.2 as advisor needs Claude Code 2.1.321 or later and Fable access.
- The advisor reads the whole conversation on every call, and an agent's history grows every turn. That read is not cached, so each call uses Fable's allowance on top of the main model (how weekly limits work). On some plans a Fable advisor bills to usage credits.
- Outside Claude Platform on AWS, it stays off when feature-flag fetching is off (for example with
DISABLE_TELEMETRYset). -
/advisor offturns it off;CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1disables the tool entirely.
Why this design works {#design}
It is the coordinator pattern turned around: instead of a strong model directing many workers, one capable worker asks a stronger reviewer only at the decision points. Most turns of a coding task are routine (read a file, run a test, make a small edit), and those do not need the strongest model. The few that set direction do. Paying for judgement only where it changes the result is the same trade-off as a second AI that reviews the first, applied inside one session. For what Fable 5.2 itself can do, see the Claude Fable 5.2 review; for the rest of the setup that makes long sessions reliable, see how to use Claude like a professional.
