Get alerts
AiGanak
Book a call

Claude Code /advisor: when a lighter model with a strong advisor saves money, and when it costs more

Key takeaways
  • The advisor lets Claude Code run a lighter main model, such as Sonnet, and consult a stronger model, such as Opus or Fable, at key moments: before committing to an approach, when an error keeps recurring, and before declaring a task done.
  • Turn it on with /advisor opus. It is saved as your default; /advisor off turns it off. It needs the Anthropic API (not Bedrock, Vertex or Foundry) and is experimental.
  • Each advisor call reads the whole conversation at the advisor model's rates, and that read is not cached. On a subscription it counts toward your usage limits.
  • In our worked example (a 60-step session growing to 140K tokens, 4 advisor calls) Sonnet with an Opus advisor cost 13% less than Opus throughout and 73% more than Sonnet alone. With a 300K-token session and 8 calls it cost 63% more than Opus throughout.
  • Rule of thumb: use the advisor on long tasks where a few decisions matter, keep sessions short with /clear, and use plain Sonnet for routine work.

Most of what a coding agent does is routine: reading files, running tests, making small edits. A few moments decide whether the task goes well: choosing the approach, getting unstuck, and checking the work before calling it done. Claude Code's advisor is built around that split. A lighter model does the routine steps, and a stronger model is consulted only at those moments. It sounds like an easy saving. It often is, but not always, and the difference comes down to how the advisor is billed.

What the advisor does

The advisor is a server-side tool. When it is on, your main model (the executor) can call a second, usually stronger model. The advisor receives the full conversation, including every tool call and result, and returns guidance that the main model applies before continuing. Anthropic's documentation says Claude tends to consult it before committing to an approach, when an error keeps recurring, and before declaring a task complete. The timing is decided by the model, not by fixed rules, and there is no setting to cap or force calls; you can ask for a consultation in your prompt, for example "consult the advisor before you continue".

Claude generally follows the advice, but if a recommended step fails or the files contradict it, it surfaces the conflict rather than following the advice blindly.

How to turn it on

  • In a session: run /advisor opus (or /advisor for a picker). The choice is saved to advisorModel in your user settings and applies to later sessions.
  • As a default: add "advisorModel": "opus" to your Claude Code settings file.
  • For one session: start with claude --advisor opus.
  • To stop: /advisor off. To disable it entirely, set CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1.

Pair it with a lighter main model, for example /model sonnet then /advisor opus. The advisor is experimental and works only through the Anthropic API: it is not available on Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform or Microsoft Foundry. It also stays off when feature-flag fetching is disabled, for example with DISABLE_TELEMETRY set.

Which models can advise which

The advisor must rank at or above your main model. The pairings Anthropic lists for current models:

Main modelAccepted advisors
Sonnet 5.5Fable, Opus 5 or later, Sonnet 5.5
Sonnet 5 or Haiku 5.5Fable, Opus 4.7 or later, Sonnet 5 or later, Haiku 5.5
Opus 5 or Opus 5.5Fable, Opus 5 or later
Fable 5.1Fable 5.1

Fable as an advisor needs Fable access, and on some plans its usage bills to usage credits after a one-time consent. Haiku 4.5 can call an advisor but cannot be one. If you pick an advisor that ranks below your main model, Claude Code does not attach it.

How it is billed

  • API billing: advisor tokens are charged at the advisor model's input and output rates, on top of your main model's usage.
  • Subscription plans: advisor usage counts toward your plan's usage limits (a Fable advisor bills to usage credits on plans where Fable does).
  • No cache for the advisor's read: each call reads the full transcript afresh. Your main model's prompt cache is not affected by turning the advisor on or off.

That last point is what decides whether you save. The main model reads its growing conversation from cache at a small fraction of the input rate. The advisor reads the same conversation at the full input rate of a more expensive model, every time it is called. The longer the session, the more each call costs.

Worked example: does it save money?

A 60-step Claude Code session. The conversation grows from 20,000 to 140,000 tokens; each step reads it from cache, adds 3,000 new tokens and writes 1,500 tokens of output. The advisor is called 4 times, spread through the session, each time reading the full conversation and writing 700 tokens of guidance. Prices are Anthropic's list prices per million tokens: Claude Opus 5.5 $4 input, $0.20 cache read, $20 output; Claude Sonnet 5.5 $2, $0.10 and $10.

SetupCost of the session
Opus 5.5 for every step$3.66
Sonnet 5.5 for every step$1.83
Sonnet 5.5 main + Opus 5.5 advisor (4 calls)$3.17 (Sonnet $1.83 + advisor $1.34)

The pairing saves 13% against running Opus throughout, and costs 73% more than Sonnet alone. What you buy with the extra is Opus-level judgement at the decision points. Now change the two things that matter most, how long the session gets and how often the advisor is called:

Saving vs Opus throughout2 advisor calls4 calls8 calls
Session grows to 60K tokens39%28%6%
Session grows to 140K tokens32%13%23% more
Session grows to 300K tokens21%7% more63% more

Short sessions with a few consultations save real money. Long sessions with frequent consultations can cost more than simply running Opus, because every call re-reads a large, uncached transcript at Opus input rates. On a subscription the same arithmetic applies to your usage limit.

When to use it, and when not

  • Use it for long, multi-step work where most steps are routine but the plan matters: large refactors, debugging an error that keeps coming back, tasks you want checked before they are declared done.
  • Keep sessions short. Write a handover note and start fresh with /clear when a task changes. It cuts the main model's cost and makes each advisor call cheaper.
  • Steer the number of calls. If the advisor is called more than the task needs, say so in your instructions, for example "only consult the advisor before major decisions".
  • Skip it for short tasks with little to plan: switch the main model with /model instead. And if every step needs the strongest model, run that model; the advisor adds cost without adding much.
  • Alternatives: opusplan uses Opus in plan mode and Sonnet for execution; subagents with their own model run a whole subtask on a chosen model.

How to tell if you need it

The signal is simple: how many of your steps ran on your most expensive model without needing it, such as approvals, status checks and short answers. If that share is high, a lighter main model with an advisor, or plain Sonnet, will stretch your limit or budget. AiGanak Scan measures this from your Claude Code logs, together with how long your sessions get, and gives you the setup to paste.

Sources

  1. Claude Code docs: Escalate hard decisions with the advisor tool (checked 16 November 2026)
  2. Claude API docs: Advisor tool
  3. Anthropic: Claude API pricing (checked 9 November 2026)

Provider prices and limits change often. Figures were checked on 16 November 2026; confirm current terms with each provider before you rely on them.

FAQ

Questions about this topic

Can't find it? Press ⌘K and ask.

Does /advisor make Claude Code cheaper?

It can. In our worked example, Sonnet 5.5 with 4 Opus 5.5 advisor calls cost 13% less than Opus throughout. But each call re-reads the whole conversation at the advisor's full input rate, so long sessions with many calls can cost more than running Opus throughout.

Does the advisor count toward my Claude subscription limit?

Yes. On subscription plans advisor usage counts toward your plan's usage limits; a Fable advisor bills to usage credits on plans where Fable usage does.

How do I turn the advisor off?

Run /advisor off, or set CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 to disable it entirely.

Keep reading

Related

AiGanak Scan: stop hitting your AI limits

AiGanak Scan is a free tool that reads your Claude, ChatGPT or Gemini chat export, or Claude Code logs, and gives you an AI Efficiency Score out of 100, how much further your plan could go, the API value of your use and the share spent re-reading long chats. It ranks ready fixes, each with a prompt to copy and steps, such as moving a long chat to a fresh one with a handover note, keeping documents in a Project, keeping tool output short and approving in one go. It is private by design: chats are never stored or shared, and no one at AiGanak sees them.

Read →

How prompt caching works on OpenAI, Anthropic and Gemini, and how much it saves

Prices checked 28 October 2026: cache reads cost 2.5% to 10% of the input rate (GPT-6 Astra $1.00, GPT-6.1 Sol $0.10, Claude Fable 5.1 $0.25, Claude Opus 5.5 $0.20, Claude Sonnet 5.5 $0.10, Gemini 4 Argon $0.10, Gemini 3.8 Flash $0.075 per million tokens). Cache writes cost 1.25x input for the short lifetime (30 minutes on OpenAI GPT-5.6 and later, 5 minutes on Anthropic), 2.5x for 12 hours on OpenAI and 2x for 1 hour on Anthropic; Gemini implicit caching has no write charge. OpenAI and Gemini cache automatically, Anthropic on request. Minimums: 1,024 tokens on OpenAI, 512 on current Claude models, 4,096 on current Gemini models. Worked example: caching a 20,000-token fixed prompt cuts cost per request by 80% to 87%.

Read →

How to reduce LLM API costs: nine levers, in the order to apply them

LLM API bills are usually dominated by input tokens. Trim system prompts, retrieve fewer documents, shorten answers, cap agent steps, use structured outputs, cache repeat answers, route simple work to smaller models, turn on provider prompt caching and batch work that can wait, applying each change on top of the last so savings are not counted twice.

Read →

Is your Claude Code setup costing more than it needs to?

Drop your Claude Code logs into AiGanak Scan. It measures how many steps ran on your top model without needing it, and gives you the exact /model and /advisor setup to fix it.