# Fable 5.2 effort levels: what they mean and how to choose

> Fable 5.2 effort levels explained: five levels, high by default in the API and medium in the apps, which to pick per task and how effort affects cost.

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

**On Claude Fable 5.2, released by Anthropic on 9 November 2026, reasoning is always on and is controlled with a single setting: effort, with five levels (low, medium, high, xhigh and max).** The default is high in the API and medium in the Claude apps and in Claude Code. There is no switch to turn thinking off: you choose how much room to give it.

This choice decides what you pay and how long you wait, more than the model name does. For small and medium businesses the method is simple: start from the recommended level, try going lower on simple tasks, and go higher only where a measurement justifies it. This guide explains how, with data from the official pages consulted on 9 November 2026. For the full model overview, see the article [Claude Fable 5.2: what it is, what changes, what it costs](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>Fable 5.2 has five effort levels: low, medium, high, xhigh and max (<a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">Anthropic documentation</a>).</li>
<li>The default is high in the API and medium in the Claude apps and in Claude Code, according to the <a href="https://platform.claude.com/docs/en/models/fable-5-2/overview">model card</a> and the <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">announcement</a>.</li>
<li>Adaptive thinking is the only mode and cannot be disabled: asking in the prompt not to reason does not work from medium upwards.</li>
<li>Anthropic recommends starting from high and moving up to xhigh only for long coding and agent work where a gain has been measured; max is for testing the model's limit.</li>
<li>In the announcement's charts, on some tests Fable 5.2 at low or medium effort matches or beats Opus 5.5 at maximum for a fraction of the cost. These are figures declared by Anthropic.</li>
</ul>
</div>

## What is adaptive thinking in Fable 5.2?

Adaptive thinking is how Fable 5.2 decides on its own how much to reason before answering: little for a simple question, a lot for a problem with several constraints. On the [official page](https://platform.claude.com/docs/en/models/fable-5-2/overview) it is listed as “always on” and as the model's only reasoning mode.

For developers, the consequence is precise. According to the [documentation](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2), the `thinking` setting with type `enabled` (with `budget_tokens`) or with type `disabled` returns a 400 error. Omit `thinking` or send `adaptive`, and control depth with `effort`. The same applies to forced `tool_choice` , which returns 400; if you are coming from Fable 5.1 or Opus 5.5, neither rule is new. The full picture of the changes is in [Migrating to Fable 5.2: API changes and 400 errors](https://giallostudio.it/en/blog/fable-5-2-api-migration).

If you only use the Claude app, none of this affects you: the level is set from the interface, where available on your plan.

## What are the five effort levels and which one is the default?

The levels are low, medium, high, xhigh and max. The [documentation](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) sets the API default to high and invites you to start there, because most of the capability gain is already present at that level. The default in the Claude apps and in Claude Code is instead medium, as stated in the [official announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2).

| Level | When to use it | Anthropic's note |
|---|---|---|
| low | Extraction, classification, triage, short answers | In the announcement's charts, on the HLE test it matches Opus 5.5's best score |
| medium | Everyday work in the apps, drafts, short analyses | Default in the Claude apps and in Claude Code |
| high | Starting point in the API; analysis and code with non-trivial logic | API default; most of the gain is already here |
| xhigh | Long coding and agent work, where an advantage has been measured | On FrontierCode 2.0 it gives the highest score among the released models Anthropic tested |
| max | Measuring the model's ceiling; rare cases where time and tokens do not matter | Used for most of the benchmarks in the announcement |

The “recommended” levels are Anthropic's reading; the pairing of each level with a type of task, in the “When to use it” column, is our own starting suggestion, to be checked against your cases.

## How much does a lower effort deliver? The figures declared by Anthropic

In the [official announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2) , Anthropic shows, for some tests, the curve between score and cost at each level. Four examples, all declared by the vendor and not replicated by third parties:

- **Programming (FrontierCode 2.0):** according to the [announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2), at high Fable 5.2 scores 45.3%, almost the same as Opus 5.5's 45.5% at maximum, for about half the cost per task. At xhigh it reaches 49.1%, the highest score among the released models Anthropic tested.
- **Agentic data gathering (WANDR):** at medium, Fable 5.2 matches Opus 5.5's best score for about 40% of the cost per task.
- **Expert knowledge (Humanity's Last Exam):** at low, it matches Opus 5.5's best score for less than a tenth of the cost per task.
- **Chart reading (Chartography):** at medium it scores 66.4%, ahead of Opus 5.5's 64.4% at maximum, for less than a tenth of the cost.

There is a flip side: where difficulty is highest, the high level is needed. On the hardware engineering test BenchCAD 2.1, at max Fable 5.2 scores 83.5 out of 100, against 65.3 for Opus 5.5 and 44.4 for Fable 5.1. The honest reading is that **the right effort depends on the task**, and that the scores in the main table (all at the highest level available unless stated otherwise) do not describe what you will get at medium. For the full picture of the tests, see [Claude Fable 5.2 benchmarks explained](https://giallostudio.it/en/blog/claude-fable-5-2-benchmarks).

## Why does Fable 5.2 think even when I tell it not to?

It is one of the known behaviors. The [what's-new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) explains that Fable 5.1 usually followed a line such as “answer directly without deliberating”, while in Anthropic's tests Fable 5.2 thinks anyway at medium and above. At high and above, the model reasons before almost every reply.

<div class="callout"><strong>Warning</strong> If easy requests are slow or expensive, the fix Anthropic points to is a lower effort for those turns, not one more sentence in the prompt. If you have prompts written for Fable 5.1 or for Opus that contain the phrase “don't think”, remove the line and try different levels.</div>

A second behavior to know about: in long, multi-step tasks the model sometimes ends its turn with a list of what remains, an offer to continue, or a stop at a common obstacle. It happens more often at low and medium, according to the same page. The remedy, explained in [Fable 5.2 known limitations and how to work around them](https://giallostudio.it/en/blog/claude-fable-5-2-limitations), is an explicit completion criterion in the instructions and, for jobs that must not stop, a higher level. How to write those instructions is covered in [Prompts for Claude Fable 5.2](https://giallostudio.it/en/blog/claude-fable-5-2-prompts); for the basics of prompt writing, see [Prompts for Claude](https://giallostudio.it/en/blog/claude-prompts).

## How do you choose the effort for a purchasing department? An example

An example, declared as such: the purchasing department of a food company handles three types of work every week with an assistant based on Fable 5.2.

1. **Sorting supplier emails** into categories (order, delay, complaint, invoice). A repetitive, short task: start from low.
2. **Comparing three quotes** for packaging with different terms, delivery times and penalties. You need to hold several constraints together: high.
3. **Checking a supply contract** of dozens of pages against internal terms and flagging mismatches. This is the rare, long job: try high and, if too many mismatches are missed, xhigh.

The method for setting the level is the same for all three tasks:

<ol class="steps">
<li><strong>Collect 20 real cases</strong> per task, anonymized, with the correct answer written beforehand.</li>
<li><strong>Run at high</strong> (Anthropic's starting point) and count the correct results.</li>
<li><strong>Go down one level</strong> on the simple tasks and check whether quality holds.</li>
<li><strong>Go up to xhigh</strong> only where high is not enough and the long task justifies the wait.</li>
<li><strong>Write the chosen level next to the prompt</strong> and retest when the model changes.</li>
</ol>

## Effort, cost and time: what to know before deciding

According to the [model card](https://platform.claude.com/docs/en/models/fable-5-2/overview), effort is the control for reasoning depth, latency and cost. The price list stays that of the model: $8 per million input tokens and $40 per million output tokens (excluding VAT for Italy), as explained in [Claude Fable 5.2 API pricing and real costs](https://giallostudio.it/en/blog/claude-fable-5-2-api-pricing). In the pages consulted on 9 November 2026, there is no price multiplier per level: what changes is the amount of work, and therefore of tokens and time, that the model spends. As a result, cost per task is measured on your own cases, not calculated on paper.

Two practical notes:

- **Effort and speed are different dials.** Fast mode (up to 2.5 times faster, at $16 and $80) speeds up generation; effort changes how much the model works beforehand. The price list is in the [official documentation](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).
- **Subagents add up too.** The documentation warns that subagents can make a session's cost grow a lot: a high effort across several parallel agents multiplies the spend.

To compare price lists across the Claude family, see [Claude: model pricing compared](https://giallostudio.it/en/blog/claude-api-pricing), and to learn how to estimate spend in general, [AI API cost: how to calculate it](https://giallostudio.it/en/blog/how-to-calculate-ai-api-costs). If you have already worked with effort on Opus, the logic is the same as described in [Claude Opus 5.5 reasoning levels](https://giallostudio.it/en/blog/claude-opus-5-5-reasoning-levels): the starting point changes, which for Opus 5.5 in the API is medium and for Fable 5.2 is high.

## What is preserved thinking and why does it matter for developers?

Preserved thinking binds reasoning blocks to the conversation that produced them. With Fable 5.2 it applies to all API accounts, and Anthropic presents it in the [announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2) as a defense against anyone trying to extract the model's reasoning by modifying past context.

For anyone with an integration, three things change, according to the [documentation](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2):

- **Editing an earlier turn, the system prompt or the tools** after a reasoning block makes the request fail with a 400 error. A beta option lets you ask to discard the invalid blocks, but Anthropic describes it as a temporary fix: discarding them can increase output tokens.
- **Switching model mid-conversation** does not carry over Fable 5.2's reasoning blocks: only Fable 5.2 and Mythos 5.2 read them. In a system with fallback models, the others discard them and the request succeeds anyway; discarded blocks are not billed.
- **The block stays in the account** that produced it or in a linked account.

If you only use the app or Claude Code, you do not need to do anything. Developers will find the changes step by step in [Migrating to Fable 5.2](https://giallostudio.it/en/blog/fable-5-2-api-migration). For systems with several models, Fable's role as an advisor is covered in [Fable 5.2 as Advisor and Orchestrator](https://giallostudio.it/en/blog/fable-5-2-advisor-orchestrator).

## Where do you set the effort?

In the API it is a request parameter, called `effort`. In the Claude apps and in Claude Code the starting level is medium. Command and menu names change: for Claude Code, the practical instructions are in [Fable 5.2 in Claude Code](https://giallostudio.it/en/blog/claude-fable-5-2-claude-code), and before putting a level into production you should always check the platform documentation.

<div class="compare">
<div class="compare-yes"><h4>When to raise the effort</h4><ul>
<li>The task is long and multi-step, and a mistake halfway costs hours.</li>
<li>The model stops before it has finished: more effort and an explicit completion criterion.</li>
<li>You have measured it: on your cases the higher level gives better results.</li>
</ul></div>
<div><h4>When to lower it</h4><ul>
<li>Requests are simple and the model takes too long.</li>
<li>Volume is high and the cost per call weighs on your margin.</li>
<li>A user is waiting in front of a frozen screen.</li>
</ul></div>
</div>

## Which mistakes should you avoid with Fable 5.2 effort?

- **Writing “don't think” in the prompt.** From medium upwards it is not enough: you need to lower the level.
- **Leaving max on out of habit.** Anthropic reserves it for measuring the ceiling and for cases where time and tokens do not matter.
- **Comparing your own results with the benchmark table.** Those scores are at the highest level available, unless stated otherwise.
- **Using the same level for everything.** Sorting emails and checking a contract have opposite needs.
- **Ignoring subagents.** A high effort across several parallel agents pushes the bill up more than the model itself.

## Who owns effort in the company, and how

The effort level is a cost item to manage just like the model. List the workflows that run on Fable 5.2, assign each one a measured level, write it next to the prompt and retest whenever Anthropic releases a new version: three versions in five months, from 9 June to 9 November 2026, make this habit less optional than it seems. Anthropic itself recommends, for most workloads, starting from Opus and moving to Fable 5.2 when tests at higher effort are not enough: to choose well between the two, see [Fable 5.2 vs Opus 5.5](https://giallostudio.it/en/blog/fable-5-2-vs-opus-5-5).

If you want a test on your own cases, with the right level for each workflow before signing a subscription or an integration, [Giallo Studio's AI consultancy](https://giallostudio.it/en/ai-consulting) starts precisely from a measurement of quality, time and cost per task. Before deciding, check the [official Anthropic pages](https://www.anthropic.com/claude-fable-and-mythos-5-2): levels, defaults and prices may change with future versions.

## Frequently asked questions

### What are the effort levels of Fable 5.2?

There are five: low, medium, high, xhigh and max. The default is high in the API and medium in the Claude apps and in Claude Code. Both the model card and Anthropic's announcement state this (consulted on 9 November 2026).

### Can you turn off reasoning in Fable 5.2?

No. Adaptive thinking is always on and cannot be disabled: in the API, a thinking setting of type enabled with budget_tokens, or of type disabled, returns a 400 error. Omit the parameter or set it to adaptive, and control reasoning depth with effort.

### Why does Fable 5.2 still think when my prompt tells it to answer right away?

Because at medium effort and above, a line like “answer directly without deliberating” does not stop it, as Anthropic found in its own tests. At high and above, the model thinks before almost every reply. For simple requests, the fix is to lower the effort for those turns, not to rewrite the prompt.

### Which effort level should I start with?

Start with high in the API, as Anthropic recommends, because most of the capability gain is already there. Then try going lower on simple tasks and going up to xhigh only on long coding and agent jobs where you have measured an advantage. Max is for cases where time and tokens do not matter.

### Does a higher effort cost more?

Generally yes, because according to the documentation effort controls reasoning depth, latency and cost. In the pages consulted, Anthropic does not publish a price multiplier per level: cost per task has to be measured on your own cases. In its charts, however, on some tests Fable 5.2 at lower levels matches Opus 5.5 at maximum while spending much less.

### What is preserved thinking and what changes for API users?

It is the mechanism that keeps reasoning blocks bound to the conversation that produced them. If you change anything that comes before them, for example a past turn, the request to Fable 5.2 returns a 400 error. With Fable 5.2 it applies to all API accounts.
