Advisor
An experimental model-tiering tool introduced by Anthropic in the Claude Code terminal. It allows the primary model to call upon a more capable advisor model at key decision points to provide strategic judgment or review opinions. The advisor model does not directly execute code.
Timeline (last 90 days)
The main model does not unconditionally adopt the advisor's opinions: if the project's existing evidence conflicts with the advisor's suggestions, the main model presents the disagreement to the user rather than blindly executing it.
In Advisor mode, the advisor model's capability level cannot be lower than the main model's; for example, when the main model is Sonnet 5.5, the advisor can only be chosen from the Fable or Opus 5 level, and smaller models such as Haiku 4.5 cannot serve as the advisor.
The advisor model itself does not write code or take over the keyboard; it only outputs strategic judgments or review opinions when invoked, and actual code execution is still performed by the main model.
There are three ways to enable Advisor in Claude Code: type `/advisor` to select interactively, type `/advisor opus` or `/advisor fable` to specify directly and write it to user settings, or add the `--advisor opus` parameter at startup to apply it only to the current session.
Claude Code allows only one main model paired with one advisor in a session; the four-model chain circulating on X — Fable 5.2 as thinker, Sonnet 5.5 as coder, Opus as reviewer, and Haiku as subagent — is not an official single-command mode.
Tokens generated by Advisor calls are billed to the account at the rate of the invoked model; the Fable series has a separate billing rule: even if the user has a subscription plan, Fable usage may be deducted directly from the usage quota, and before first enabling it the terminal requires a one-time confirmation of quota authorization.
The community claims the Advisor feature reduces costs by 3.4x, but that figure is not in Anthropic's official documentation; officially, it is only qualitatively confirmed that invoking a strong model at decision points rather than using an expensive model throughout can reduce costs.
When the advisor is invoked, it re-reads the full session as input, and this reading does not enter prompt caching, so each summon is billed at full price for input tokens; the main model's cache is unaffected, and the advisor's returned suggestions are cached locally in subsequent steps.
In the organization, each supervisor uses the Opus 5.5 model and employees use the Sonnet 5.5 model; supervisors can enable the Advisor feature and consult Fable 5.1 when they run into difficulties.
All Facts (8)
In Advisor mode, the advisor model's capability level cannot be lower than the main model's; for example, when the main model is Sonnet 5.5, the advisor can only be chosen from the Fable or Opus 5 level, and smaller models such as Haiku 4.5 cannot serve as the advisor.
50%UnverifiedThe advisor model itself does not write code or take over the keyboard; it only outputs strategic judgments or review opinions when invoked, and actual code execution is still performed by the main model.
50%UnverifiedThere are three ways to enable Advisor in Claude Code: type `/advisor` to select interactively, type `/advisor opus` or `/advisor fable` to specify directly and write it to user settings, or add the `--advisor opus` parameter at startup to apply it only to the current session.
50%UnverifiedClaude Code allows only one main model paired with one advisor in a session; the four-model chain circulating on X — Fable 5.2 as thinker, Sonnet 5.5 as coder, Opus as reviewer, and Haiku as subagent — is not an official single-command mode.
50%UnverifiedTokens generated by Advisor calls are billed to the account at the rate of the invoked model; the Fable series has a separate billing rule: even if the user has a subscription plan, Fable usage may be deducted directly from the usage quota, and before first enabling it the terminal requires a one-time confirmation of quota authorization.
50%UnverifiedThe community claims the Advisor feature reduces costs by 3.4x, but that figure is not in Anthropic's official documentation; officially, it is only qualitatively confirmed that invoking a strong model at decision points rather than using an expensive model throughout can reduce costs.
50%UnverifiedWhen the advisor is invoked, it re-reads the full session as input, and this reading does not enter prompt caching, so each summon is billed at full price for input tokens; the main model's cache is unaffected, and the advisor's returned suggestions are cached locally in subsequent steps.
50%UnverifiedThe main model does not unconditionally adopt the advisor's opinions: if the project's existing evidence conflicts with the advisor's suggestions, the main model presents the disagreement to the user rather than blindly executing it.
50%