# Claude Advisor Orchestrator: How to Use Fable 5.2 in AI Agents

> Claude Advisor Orchestrator: Fable 5.2 can advise smaller models or coordinate them. Rules, costs and an SME example, from Anthropic's docs.

Source: Giallo Studio (giallostudio.it) · Page: https://giallostudio.it/en/blog/fable-5-2-advisor-orchestrator · Published: 2026-11-09 · Updated: 2026-11-09

**Claude Advisor Orchestrator** refers to the two roles that Anthropic described on 9 November 2026 for Claude Fable 5.2 in AI agent systems. In the role of **Advisor**, Fable 5.2 is the consultant that a smaller model (Haiku, Sonnet or Opus) turns to when it faces a hard problem. In the role of **Orchestrator**, it is the coordinator that delegates simpler tasks to cheaper models. This comes from the [Anthropic official announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2)says, as checked on 9 November 2026.

The idea is practical: the most expensive model works only where it is needed. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, against $4 and $20 for Opus 5.5 and $2 and $10 for Sonnet 5.5 ([official page](https://platform.claude.com/docs/en/models/fable-5-2/overview)). If you use it for every step of a long workflow, the bill goes up. If you call it only on the hard nodes, spending stays close to that of a mid-range model.

This article explains how the two roles work, which technical rules the documentation states, how much the cost can change (with a transparent worked example) and how to set up an architecture for an SME, with human approvals. For the full model sheet, see [Claude Fable 5.2: what it is and what changes](https://giallostudio.it/en/blog/claude-fable-5-2-what-it-is).

<div class="key-points">
<p class="kp-title">In summary</p>
<ul>
<li>Advisor and Orchestrator are the two roles of Fable 5.2 in agents that Anthropic lists in its <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">announcement of 9 November 2026</a>: in the first, Fable advises a smaller model; in the second, it coordinates it.</li>
<li>If the executor is Fable 5.2, the advisor must be Fable 5.2 or Mythos 5.2: any other advisor returns a 400 error, according to the <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">official documentation</a>.</li>
<li>The advisor tool is in beta; a Fable 5.2 advisor returns its advice encrypted, so any code that reads the advice text needs to be checked (<a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">source</a>).</li>
<li>Subagents can account for a large part of what a session costs, warns the <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">same page</a>: savings are not automatic and have to be measured.</li>
<li>Anthropic recommends starting with Opus 5.5 and moving up to Fable only if tests are not enough, so a single cheap model is often the best choice.</li>
<li>Whatever the architecture, anything with consequences (payments, sends, price list changes) goes through human approval.</li>
</ul>
</div>

## What are Advisor and Orchestrator in Fable 5.2?

They are two ways of combining several models in the same job, pointing in opposite directions. The difference is who is in charge.

In the **Advisor** role, control stays with a small model, the executor. It carries out the task step by step and, when it meets a hard node, asks Fable 5.2 for an opinion. Think of a junior colleague who knows when to raise a hand.

In the **Orchestrator** role, control passes to Fable 5.2. It splits the work, assigns the simple parts to smaller models (the “subagents”), checks the results and puts them back together. Think of the site foreman.

| Role | Who works on the task | When Fable 5.2 steps in | What you pay Fable for |
|---|---|---|---|
| Advisor | Haiku, Sonnet or Opus | Only when the executor asks for help | Only the consultations |
| Orchestrator | Fable 5.2 coordinates, smaller models execute | At every planning and checking step | Planning and verification, not execution |
| Fable only | Fable 5.2 | Always | All the work |
| Mid-range model only | Opus 5.5 or Sonnet 5.5 | Never | None |

The first two rows follow the wording of the announcement. The “What you pay Fable for” column is our own reading: Anthropic does not publish a cost breakdown by role.

The official context is this: Fable and Mythos, Anthropic writes, are designed for the most ambitious, open-ended technical work, while for most coding and knowledge tasks you start from Opus 5.5, moving down to Sonnet 5.5 or Haiku 5.5, or up to Fable, depending on difficulty. Advisor and Orchestrator are a way to combine the tiers instead of picking only one.

## What technical rules apply to the Advisor?

According to the [“What's new” page of the documentation](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2), the **advisor tool** is in beta, and one rule applies: the advisor must be at least as capable as the executor. Three consequences follow.

<ul class="checklist">
<li>With a Fable 5.2 or Mythos 5.2 executor, the advisor can only be Fable 5.2 or Mythos 5.2. Any other, including Fable 5.1 and Mythos 5.1, returns an <code>invalid_request_error</code> with code 400.</li>
<li>If you already have an executor and only want to change the advisor, set <code>claude-fable-5-2</code> as the advisor's model: the documentation says every executor model accepts it.</li>
<li>Changing the advisor's model is the same as changing the list of tools (<code>tools</code>): the documentation asks you to do it in new conversations, because sending earlier thinking blocks back after the change fails the prefix check with a 400 error.</li>
</ul>

There is one detail developers should not overlook: a Fable 5.2 advisor returns its advice **encrypted**. If your code reads the advice text (to write it to a log, for example), check the documentation section on result variants. All the upgrade steps are in the guide [Migrating to Fable 5.2: API changes and 400 errors](https://giallostudio.it/en/blog/fable-5-2-api-migration). We do not list the advisor tool's parameter names here beyond those quoted: you will find them in the documentation, and it is best to copy them from there.

## How does the Orchestrator work, and why do subagents cost money?

The announcement gives two Anthropic examples of how Fable 5.2 coordinates other agents. They are vendor statements, not verified by third parties.

- In one experiment, an engineer asked Fable 5.2 to design a prototype operating system: starting from an empty folder, the model coordinated more than 1,000 subagents, working for eight days, and produced a working system that boots on a laptop.
- A Datadog tester, quoted in the same announcement, describes the model as “almost like an incident commander that can send out faster models to get quick signals and coordinate”. Another, from Uber, describes a 10–12 hour security audit with about 170 reviewer agents.

These are extreme cases, far from an SME. They help to understand the mechanism: the large model plans, the small ones execute in parallel.

The documentation puts the other side of the coin among the behaviors to know about: how much Fable 5.2 delegates to subagents depends on the system hosting it and on the tasks; in an agentic coding product whose prompts invite delegation, it delegates readily, and **subagents can account for a large part of what a session costs**. So delegating does not mean saving. It means the cost shifts and has to be measured. Limits of this kind, with remedies, are collected in [Fable 5.2 known limitations and how to work around them](https://giallostudio.it/en/blog/claude-fable-5-2-limitations).

## How much do you save with Fable as Advisor?

It depends on how often the executor asks for help. Here is a worked example, with made-up but declared numbers, calculated on the prices of the [official pricing](https://platform.claude.com/docs/en/models/fable-5-2/overview): 1,000 tasks, each with 20,000 input tokens and 2,000 output tokens.

| Scenario | Cost per task ($) | Cost per 1,000 tasks ($) |
|---|---|---|
| Sonnet 5.5 only (2 / 10 per million) | 0,06 | 60 |
| Fable 5.2 only (8 / 40 per million) | 0,24 | 240 |
| Sonnet 5.5 with Fable as advisor on 100 tasks out of 1,000 | 0.072 on average | 72 |

The last scenario, with the prices from the [price list](https://platform.claude.com/docs/en/models/fable-5-2/overview), assumes that in 100 tasks out of 1,000 Fable receives 10,000 input tokens and produces 1,000 output tokens: that is $0.08 plus $0.04, so $0.12 per consultation, $12 in total. The result is a simplified assumption: how the advisor tool actually counts the tokens of the consultation has to be checked in the documentation, and Fable's adaptive thinking (always on) adds tokens that this calculation does not include. The point is not the exact figure. It is the order of magnitude: with a cheap executor and rare consultations, you pay for Fable as a second opinion, not as the engine. More worked examples are in [Claude Fable 5.2 API pricing and real costs](https://giallostudio.it/en/blog/claude-fable-5-2-api-pricing) and in the guide [how to calculate the cost of an AI API](https://giallostudio.it/en/blog/how-to-calculate-ai-api-costs).

<div class="callout"><strong>Warning</strong> A saving on price per token is not a saving on the result. If the cheap executor makes mistakes silently without asking for help, no advisor will save it. Measure quality on your own cases, not just cost: the benchmark scores Anthropic publishes are self-reported by the vendor and, as of 9 November 2026, do not appear to have been replicated by third parties in the pages we read.</div>

## Example: an auto parts distributor in Emilia

This is an example, declared as such: an auto parts distributor in Emilia receives quote requests by email and PDF every day, often with ambiguous codes, brand abbreviations and photos of parts. It wants an agent that prepares draft offers.

One possible architecture using the roles we are discussing:

<ol class="steps">
<li><strong>Reading and sorting.</strong> A small model (Haiku or Sonnet) reads the request, extracts codes and quantities and checks that no data is missing.</li>
<li><strong>Catalog search.</strong> The same executor looks up the parts in the management system using read-only tools.</li>
<li><strong>Doubtful case, consult the advisor.</strong> When two codes are only partly compatible, or the request is contradictory, the executor asks Fable 5.2 for its opinion as advisor.</li>
<li><strong>Draft offer.</strong> The agent prepares the draft without sending it.</li>
<li><strong>Human approval.</strong> A salesperson checks the draft and the price and sends it themselves.</li>
</ol>

Fable costs money only in step 3, which is the one where a mistake loses margin. Write operations (sending to the customer, changing price lists) stay out of the agent's reach. If the workflow grows, with more parallel agents, you move to an orchestrator; in that case it is best to give each subagent only the tools it needs, as explained in [Claude Managed Agents: network, web fetch and security](https://giallostudio.it/en/blog/claude-managed-agents-network-web-fetch).

## Advisor or Orchestrator: which should you choose?

<div class="compare">
<div class="compare-yes"><h4>Choose the Advisor if</h4><ul>
<li>The workflow is linear and almost always easy, with only a few really hard nodes.</li>
<li>You want to keep the cost close to that of a mid-range model.</li>
<li>You already have a working agent on Sonnet or Opus and want to raise its reliability only where it goes wrong.</li>
</ul></div>
<div><h4>Choose the Orchestrator if</h4><ul>
<li>The work splits into many independent parts that can run in parallel.</li>
<li>You need a model that plans, checks the results and puts the pieces back together.</li>
<li>You are willing to measure the cost of subagents, which can become the main item.</li>
</ul></div>
</div>

Our advice, as consultants: start with the Advisor. It is easier to govern, easier to measure and the cost is more predictable. The Orchestrator makes sense when the work is truly parallel. To choose the executor's model, see [The best AI model for agents](https://giallostudio.it/en/blog/best-ai-model-for-agents); for the basic concepts, see [AI agents: what they are and how they work](https://giallostudio.it/en/blog/what-are-ai-agents).

## When do you need neither?

Often. The [what's-new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) says to start from the latest Opus model and move to Fable 5.2 only when tests on Opus at a higher effort are not enough. If a single model, Opus 5.5 or Sonnet 5.5, solves the task at the quality you need, adding a second tier adds complexity, possible errors and one more log to check. The comparison with the “task → model” matrix is in [Fable 5.2 vs Opus 5.5](https://giallostudio.it/en/blog/fable-5-2-vs-opus-5-5).

There is also a data issue. Fable 5.2 is a “Covered Model”: according to the [Anthropic page](https://support.claude.com/en/articles/15425695-covered-models), read on 9 November 2026, it requires retention of at least 30 days. We infer, though it is not stated explicitly, that this also applies to what you send to Fable as advisor: so do not put data into the consultation that you do not want retained. We discuss this in [Fable 5.2 and privacy](https://giallostudio.it/en/blog/fable-5-2-privacy-data-retention).

## Which mistakes should you avoid?

- **Using Fable as the executor “to be safe”.** That loses the economic advantage of the whole architecture.
- **Leaving delegation unrestricted.** Without limits on the number of subagents and on run time, the cost of a session can grow unexpectedly.
- **Not logging the consultations.** If you do not know how often the executor calls the advisor and with what outcome, you can neither improve nor control spending.
- **Giving the agent write tools.** An agent that can send emails or change price lists must not be able to do so without approval.
- **Changing the advisor mid-conversation.** It is a change to `tools`: sending earlier thinking blocks back after the change produces a 400 error, so the change is made in new conversations.

## How do you get started in your business?

<ul class="checklist">
<li>Choose one process with many easy cases and few hard ones.</li>
<li>Collect 20 real examples with the right answer already known.</li>
<li>Measure the executor on its own: quality, time, cost per task.</li>
<li>Add the advisor and measure again. If quality does not rise, it is not worth the spend.</li>
<li>Put human approval on everything that leaves the company.</li>
</ul>

At Giallo, this is how we work with SMEs: one process, a sample of real cases, numbers before scaling. If you want to check whether an agent with an advisor makes sense for your workflow, [Giallo Studio designs custom AI agents](https://giallostudio.it/en/ai-agents) with human approvals and costs measured on your own cases, not on the vendor's tables.

## Frequently asked questions

### What is the Advisor role of Claude Fable 5.2?

It is a role in which a smaller model, such as Haiku, Sonnet or Opus, works on the task and consults Fable 5.2 when it hits a harder problem. Anthropic describes it in its announcement of 9 November 2026. The large model steps in only on the steps that need it, not on every request.

### What is the Orchestrator role?

It is the opposite role: Fable 5.2 leads the work and delegates simpler tasks to smaller models. Anthropic lists it in the announcement together with the Advisor. The large model plans and checks, while the cheaper ones execute.

### Which model can I use as advisor if the executor is Fable 5.2?

Only Fable 5.2 or Mythos 5.2. The documentation states that the advisor must be at least as capable as the executor, and any other advisor, including Fable 5.1 and Mythos 5.1, returns a 400 error. The advisor tool is in beta.

### Should I always use Fable 5.2 as advisor?

No. Anthropic recommends starting with Opus 5.5 for most jobs and moving up to Fable only when tests on Opus at a higher effort are not enough. For simple, repetitive tasks, a single cheap model with no advisor is often the most sensible choice.

### Do orchestrators make you spend more?

They can. Anthropic's documentation warns that subagents can account for a large part of what a session costs, and that how much the model delegates depends on the system hosting it and on the tasks. Measure the cost per task on a real sample before putting the agent into production.

### Do you need human oversight in an architecture with advisor and orchestrator?

Yes, on the steps with consequences: payments, messages sent to customers, changes to data and price lists. A more capable model makes fewer mistakes, but it is not infallible, and benchmark scores are self-reported by the vendor. The model proposes, a person approves.
