Advisor Model (escalate hard decisions)

Key Principle

The advisor tool pairs the main model with a stronger advisor model that Claude consults at decision points — before committing to an approach, when stuck on a recurring error, or before declaring a task complete (Section: Advisor tool; p. advisor). It runs server-side on Anthropic infra, receives the full conversation including every tool call and result, and returns guidance Claude applies before continuing. It lets you run a cheaper/faster main model and pay for top-tier capability only when it matters.

Why This Matters

Without it you either over-pay (strongest main model always on) or risk weak judgment at the moments that decide outcomes (plan quality on large refactors, recurring-error debugging, completion checks). The context-management angle: the advisor injects a stronger model's judgment without polluting the main context — its read of the conversation feeds back as guidance, not as bulk content the main model must carry. Experimental; behavior, pricing, and availability may change.

Good Examples

Enable (three ways):

/advisor              # picker
/advisor opus         # direct; saves advisorModel in user settings, persists across sessions
claude --advisor opus # single session; takes precedence over advisorModel

Persistent default in user settings:

{ "advisorModel": "opus" }

Turn off / disable:

/advisor off          # stops use + clears saved advisorModel
CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1   # disables tool entirely, incl. command + flag

Steer when it's consulted (model-driven, no setting to cap or force calls):

consult the advisor before you continue

Press Ctrl+O to expand the Advising line and read the guidance.

Counterpoints

• Pairing rule: advisor ≥ main, including version. If the advisor is less capable than main, it's not attached (shown in /advisor output + a notification). Aliases opus, sonnet, fable resolve to latest; full IDs like claude-opus-4-8 accepted (Section: Choosing an advisor).
• Haiku 4.5 can call an advisor but cannot be one (accepts Fable/Opus/Sonnet). Opus 4.7 main + Opus 4.6 advisor is rejected (advisor must be at/above main's version). Fable 5 main accepts Fable only — Opus/Sonnet advisor rejected (Fable is higher tier).
• Subagents inherit the configured advisor and re-run the pairing check against their own model — a subagent may use the advisor even when the lead can't, or vice versa.
• Cost: each call sends the full conversation at the advisor's rates (API → advisor input/output rates; subscription → counts toward plan limits and /usage totals). Toggling /advisor mid-session does not invalidate the main model's prompt cache (unlike changing model/effort), but the advisor's own read of the conversation is never cached — each call reprocesses the full transcript.
• NOT available on Amazon Bedrock, Google Vertex AI, or Microsoft Foundry — Anthropic API only. Via an LLM gateway (ANTHROPIC_BASE_URL) works only if the gateway forwards the request intact to the Anthropic API.
• When NOT to use: short tasks with little to plan, or work where every turn needs the strongest model (switch the main model with /model, or use opusplan / subagents instead).

Key Commands & Config

Slash:    /advisor   |   /advisor opus   |   /advisor off   |   /model   |   /usage
Flag:     claude --advisor opus   (errors out if main model unsupported)
Setting:  "advisorModel": "opus"   (saved even if current main unsupported — activates on switch)
Env:      CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 ,  ANTHROPIC_BASE_URL
Keys:     Ctrl+O  (expand the "Advising" line)
Aliases:  opus | sonnet | fable   (also full IDs, e.g. claude-opus-4-8)

Requirements (all): Claude Code v2.1.98+ (claude update); Anthropic API only; supported main model Opus 4.6+, Sonnet 4.6, Haiku 4.5, or Fable 5 (v2.1.170+ + Fable 5 access).

Comparison with related features:

Rules of Thumb

• Reach for the advisor on long, multi-step tasks where most turns are routine but plan quality decides the outcome.
• Common pairings: Sonnet+Opus (routine, escalate planning); Haiku+Opus (lowest-cost main, strong planning); Opus+Opus (independent high-stakes check); Fable+Fable (max capability, v2.1.170+); Sonnet+Sonnet (cheap second opinion).
• If the picker shows the advisor as not attached, check the pairing rule (advisor ≥ main, same/higher version).
• Prefer the advisor over upgrading the main model when only some turns need top-tier judgment.

Related References

• Agent Teams (experimental) - subagents/teammates and how advisor inheritance interacts
• Background Sessions & Agent View (claude agents) - per-session model selection
• Claude Code Core Framework - managing finite context, scaling cost/capability
• Claude Code Rules of Thumb - when to escalate model capability