Fable 5.1 as an advisor adds just 1.7 points to Opus 5.5 at twice the cost. So when does /advisor pay off?

+1.7 points for about 2.1 times the cost. That is what Anthropic itself measured for the pair people are asking about most right now: Opus 5.5 writes the code, Fable 5.1 advises. On X the combination is sold as a way to "save thousands". Anthropic's documentation has no such figures, and for this pair its measurement points the other way.

The idea itself is good, and there are pairs where it works. Below: how /advisor works, what exactly Anthropic measured, and how to tell whether an advisor will pay off on your work.

What /advisor does in Claude Code

The main model (Anthropic calls it the executor) works as usual, but at key moments it can ask a second model for advice. The advisor gets the entire conversation: every tool call, every result. It returns guidance and the executor applies it. If a file or a command result contradicts the advice, Claude surfaces the conflict instead of following the advice blindly.

The advisor runs on Anthropic's side as a server tool and is available on both subscriptions and the API. You pick the advisor model; Claude picks the moment. According to the documentation, it most often consults in three places:

1. 1

   Before committing to an approach

   While the plan is not yet in motion and a mistake in it costs the most.

2. 2

   When an error keeps recurring

   The executor is stuck on the same thing, and a fresh look is worth more than another attempt.

3. 3

   Before declaring the task done

   A check before the work is reported as finished.

The model still chooses the timing. There is no setting to force the advisor or to limit it; the only lever is wording in your prompt, for example consult the advisor before you continue.

The idea is not new. Anthropic described the advisor strategy on its blog on 9 April 2026, and the experimental Advisor Tool landed in Claude Code 2.1.117 (22 April). It is back in the conversation because of Opus 5.5: on 22 September it became the default Opus model in Claude Code, and the question "should I give it Fable 5.1 as an advisor" came up on its own.

How to turn it on and what you will see

/advisor            # pick a model from the list
/advisor fable      # Fable right away
/advisor off        # turn it off
claude --advisor opus   # this session only

The choice is saved in your user settings as advisorModel. After a session starts you get a notice that the Advisor Tool is on and may use more tokens. When the executor consults, the transcript shows an Advising line with the advisor's model name, and Ctrl+O opens the advice text. Subagents inherit the advisor.

The main constraint: the advisor cannot be weaker than the executor. For Opus 5 and Opus 5.5 only Fable and Opus 5+ qualify. If a pair does not pass, Claude Code does not attach the advisor and warns you.

Fable comes with a catch. On some plans Fable is billed from usage credits rather than from your subscription limits. If your account needs a one-time consent, /advisor fable will not save the choice and sends you to /model fable instead: accept there, switch the main model back, and pick Fable as the advisor again.

What Anthropic measured for Opus 5.5 + Fable 5.1

In its guide to trading off cost and intelligence, Anthropic reports a run on an internal agentic coding benchmark. It is a plain API agent, not Claude Code, with five attempts per task.

Anthropic itself calls the 1.7-point gap "at the edge of run-to-run noise". Against Opus 5.5 at medium, its default, the pair gains 3.5 points for about 3.5 times the cost.

The most useful observation sits right there: the pair lands roughly on Opus 5.5's own effort curve. Opus 5.5 alone at xhigh scored 91.1% for $4.11 per attempt. In other words, the advisor buys about what letting the model think longer buys.

Two more details explain where the money goes. Opus 5.5 asked for advice about 1.4 times per attempt and skipped it entirely in 4% of attempts. The executor's own spend barely moved ($1.36 against $1.38 without an advisor), while the consultations cost $1.55. There is no saving on the executor; the advisor comes on top.

So where do the savings stories come from

From earlier measurements on other pairs. In the April blog post Anthropic showed Sonnet with an Opus advisor on SWE-bench Multilingual: +2.7 points over Sonnet alone and 11.9% lower cost per task. Haiku with an Opus advisor scored 41.2% on BrowseComp against 19.7% for Haiku alone.

These pairs share one thing: the executor is clearly weaker than the advisor. Anthropic puts it plainly: the advisor can only hand over capability the executor lacks. When the executor is already close to the advisor, you pay and there is little to hand over. For reference, on a SWE-bench Pro subset Opus 5.5 at default matched Fable 5.1 at default for about a fifth of the price per solved task. The gap between them is small, and the measurement above confirms it. We compared Fable 5.1 with its competitors separately.

Why the advisor is not cheap

The advisor has no cache. Each call reads the full transcript again at the advisor's rate, with nothing reused between calls. Prices as of 5 October 2026, per million input/output tokens: Opus 5.5 $4/$20, Fable 5.1 $10/$50. In a long session where the transcript keeps growing, each new piece of advice costs more than the last.

On a subscription the spend counts against your plan limits, and on some plans a Fable advisor draws from usage credits. Advisor spend shows up in /usage. If you already run into the weekly limit, an advisor gets you there sooner; we worked out how context length eats the limit on the September numbers.

When /advisor pays off

1. 1

   The executor is clearly weaker than the advisor

   Sonnet or Haiku with an Opus or Fable advisor. These are the pairs the documentation suggests: routine work on a fast model, decisions with a strong one.

2. 2

   The task is long and multi-step

   A large refactor, debugging a recurring error, a check before handing work over. On short tasks the benefit is small.

3. 3

   The executor actually asks

   Count how often Advising appears in the transcript. Almost never means you get a plain run; on every task means running the advisor's model directly is cheaper.

4. 4

   For Opus 5.5, compare against effort

   Before calling in Fable, raise Opus 5.5 to xhigh on the same tasks. In Anthropic's measurement that gets you about the same result.

Our opinion, open to argument: with Opus 5.5 there is no reason to keep a Fable 5.1 advisor on all the time. It makes sense as a targeted move on one hard task where Opus has already failed twice. We are wrong if, on your tasks, the gap between Opus 5.5 and Fable 5.1 is wider than on Anthropic's benchmark. That takes an evening to check: 10 to 20 typical tasks, two runs, and a look at /usage.

An advisor works well in a "fast model plus strong model" pair. For two strong models it buys about what extra thinking time buys, only at a higher price. If you want to see how to count the price of a task rather than a token, read our GPT-6 Sol vs Opus 5.5 breakdown.

Versions and prices as of 5 October 2026: Claude Code with the experimental Advisor Tool, Opus 5.5, Fable 5.1. Versions and prices change, so check the current ones.