# Claude computer use with Fable 5.2: letting the model operate a PC on its own

> Claude computer use with Fable 5.2: 89.3% on OSWorld 2.1, the new computer_toolset_20260801 and the 400 error on the old tool. What to automate, and how safely.

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

**Claude computer use is Claude's ability to work in front of a screen the way a person would, even on software with no API. With Claude Fable 5.2, released by Anthropic on 9 November 2026, it runs through a new toolset, `computer_toolset_20260801`.** In the [official announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2) table, Fable 5.2 scores 89.3% on OSWorld 2.1, the computer-use test, ahead of Opus 5.5 at 87.2% and Fable 5.1 at 83.5%. These figures are reported by Anthropic.

For an Italian office the promise is clear: the management software from the 2000s, with no API, where someone today has to copy data from one window to another. The reality is more sober. The model has improved, but the environment, the permissions and the human oversight still have to be built. This guide explains what has changed, what you can hand over to computer use and what you cannot. The general overview of the model is in the [guide to Claude Fable 5.2](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 and Mythos 5.2 support computer use through the <code>computer_toolset_20260801</code>toolset, according to the <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">official documentation</a> accessed on 9 November 2026.</li>
<li>A request that declares the old <code>computer_20251124</code> gets a 400 error on every platform except Amazon Bedrock, <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">as stated by Anthropic</a>.</li>
<li>On OSWorld 2.1, Fable 5.2 scores 89.3%, Opus 5.5 87.2% and Fable 5.1 83.5%: <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">figures reported by Anthropic</a>, on a partial version of the test.</li>
<li>Computer use is for where an API is missing: dated management systems, supplier portals, desktop programs. Where an API exists, it remains the better route.</li>
<li>An isolated environment, minimal permissions and human approval of actions with consequences are not optional: they are the project.</li>
</ul>
</div>

## What is Claude computer use and how does it work?

Computer use lets a model act on a graphical interface. The model receives what is on the screen, decides the next step (open an item, click a field, type a value) and proposes it. The program hosting it carries out the step, returns a new image of the screen and the loop restarts until the task is done.

The part that is often overlooked is this: the model is not a robot that walks into your PC on its own. Someone has to set up a dedicated machine or browser, connect the tools and decide what the model may do. The practical side for developers, with classes and tool loops, is already covered in [Claude SDK: browser and computer use to automate a management system](https://giallostudio.it/en/blog/claude-sdk-browser-computer-use). Here we focus on what Fable 5.2 brings.

## What changes with Fable 5.2: the new computer_toolset_20260801

From the [what's-new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2), consulted on 9 November 2026, this is the picture.

| Model | Toolset `computer_toolset_20260801` | Old tool `computer_20251124` |
|---|---|---|
| Fable 5.1 | Accepted | Accepted with the beta header `computer-use-2025-11-24` |
| Fable 5.2 and Mythos 5.2 | Accepted | 400 error `invalid_request_error`, except on Amazon Bedrock |
| Fable 5.2 and Mythos 5.2 on Amazon Bedrock | Accepted | Accepted, as for Fable 5.1 |

In plain words: if your integration declares the old tool and you move to Fable 5.2, requests fail. This is not a malfunction of the model, it is a platform rule. The documentation says to declare `computer_toolset_20260801` and points to a dedicated migration guide from `computer_20251124`. We do not go into the parameters of the new toolset: they are not in the pages we consulted and have to be read in the official computer use documentation, which is linked from the [what's-new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

This change is part of a package of changes that can break code written for Fable 5.1. The full list, with the fix for each, 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 and have no integration of your own, this section does not apply to you.

## Does a high OSWorld 2.1 score mean the model works on its own?

No. OSWorld 2.1 measures whether a model completes tasks on a computer in test environments. In the [Anthropic's announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2) the row is labeled “offline set, partial”: part of the test, run on offline sets. The scores are reported by the vendor and, as of 9 November 2026, we found no independent verification in the pages we consulted.

| Model | OSWorld 2.1 (% of tasks completed) |
|---|---|
| Fable 5.2 | 89,3 |
| Opus 5.5 | 87,2 |
| Fable 5.1 | 83,5 |
| Gemini 4 Carbon | 81,5 |
| GPT-6 Astra | 79,5 |

The useful reading is a different one. The gap between Fable 5.2 and Opus 5.5 is 2.1 points: small, for a model that costs twice as much per token ($8/$40 against $4/$20 per million tokens, according to the [official page](https://platform.claude.com/docs/en/models/fable-5-2/overview)). The clearest jump is over Fable 5.1, at 5.8 points. If you already use Opus 5.5 for computer use, switching is not automatic: test it on your own software. The picture from the other tests is in [Claude Fable 5.2 benchmarks explained](https://giallostudio.it/en/blog/claude-fable-5-2-benchmarks), and the differences from the previous version are in [Fable 5.2 vs Fable 5.1](https://giallostudio.it/en/blog/fable-5-2-vs-fable-5-1).

A high score on test tasks is also not the success rate you will get. Your software has warning windows, custom menus and required fields that are not obvious. The real test is twenty operations from your own office, repeated.

## Which office processes suit computer use?

The criterion is simple: it pays off where there is a repeated task, on an interface that offers no better alternative, with an error that can be caught before it causes damage.

<div class="compare">
<div class="compare-yes"><h4>Works well</h4><ul>
<li>Reading data from management software or a portal with no API and copying it into a spreadsheet or another records system.</li>
<li>Preparing drafts (orders, records, entries) that a person checks before saving.</li>
<li>Downloading documents from supplier or public-body portals, with dedicated, read-only credentials.</li>
<li>Spot checks: comparing a screen with a document and flagging the differences.</li>
</ul></div>
<div><h4>Avoid, or keep under tight control</h4><ul>
<li>Payments, bank transfers, submissions to public bodies, invoice issuing: actions whose effects are hard to undo.</li>
<li>Deletions and bulk changes to records.</li>
<li>Programs with 2FA or confirmation requests handled by a person.</li>
<li>Processes with a stable API or a reliable export: there, computer use only adds fragility.</li>
</ul></div>
</div>

If your doubt is between this approach and traditional automation, the comparison with fixed-rule procedures is in [RPA vs AI automation](https://giallostudio.it/en/blog/rpa-vs-ai-automation), while if you are starting from the practical problem, see [automate a management software without API](https://giallostudio.it/en/blog/automate-business-software-without-api).

## An administrative office with old management software and no API

An example, presented as such: the administrative office of a small company uses software installed on a PC, with no API and no useful export. Every morning a person opens the suppliers' delivery notes as PDFs, looks up the matching order in the program and checks that quantities and prices match.

This is how we would set it up, in a trial scenario:

<ol class="steps">
<li><strong>Separate environment.</strong> A virtual machine with the software and a copy of the data, not the office PC with production data.</li>
<li><strong>Dedicated read-only user.</strong> The model uses an account created for it, which can see the orders but cannot change them.</li>
<li><strong>Limited task.</strong> For each delivery note the model looks up the order, reads quantities and prices and prepares an outcome line: matches, does not match, not found.</li>
<li><strong>A person in the loop.</strong> The outcome goes to the administrative staff, who check the flagged lines and decide. Nothing is written to the management software.</li>
<li><strong>Measure before extending.</strong> Twenty days already closed, for which you know the correct result: errors, time, cost per day.</li>
</ol>

Only if step 5 holds up should you consider giving the model more consequential actions, such as preparing a draft entry, always subject to approval. We do not promise savings: they depend on your volumes and must be measured in a pilot, not estimated in advance.

## What are the risks of computer use, and how do you contain them?

An agent acting on a screen can go wrong in ways that do not exist with an API. It can click the wrong button, misread a value, stop at an unexpected warning. The Fable 5.2 documentation also notes that on long tasks the model [may stop before it has finished](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2), especially at low or medium effort: in computer use, where every task is a chain of steps, that is a limit to measure. The countermeasures are concrete.

<ul class="checklist">
<li><strong>Isolated environment:</strong> a virtual machine or dedicated browser, with no access to the rest of the company network and with restricted outbound traffic.</li>
<li><strong>Minimal permissions:</strong> an account created for the purpose, read-only until something more is needed, never the owner's.</li>
<li><strong>Human approval:</strong> the approver must see what will be done (which record, which value), not a bare “proceed?”.</li>
<li><strong>No plaintext secrets:</strong> passwords and codes do not go in the instructions to the model. The environment manages the credentials.</li>
<li><strong>Action log:</strong> every step saved with a screenshot and date, so you can reconstruct what happened.</li>
<li><strong>Distrust content:</strong> a web page or a PDF can contain text that looks like an instruction. The model reads what it finds, and content must not be treated as an order.</li>
</ul>

Then there is the data question. Fable 5.2 is a “Covered Model”: according to the [Anthropic page](https://support.claude.com/en/articles/15425695-covered-models), consulted on 9 November 2026, prompts and responses are retained for at least 30 days and zero data retention (ZDR) is not available unless authorized. In computer use the prompts include screenshots: if they show customer records or special-category data, that data ends up within scope. The topic is covered in [Fable 5.2 and privacy: data retained for 30 days](https://giallostudio.it/en/blog/fable-5-2-privacy-data-retention). For network isolation and the permissions of a hosted agent, see also [Claude Managed Agents: network, web fetch and security](https://giallostudio.it/en/blog/claude-managed-agents-network-web-fetch).

## When is Fable 5.2 worth it, and when is a smaller model enough?

Anthropic itself, in the [what's-new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2), suggests starting from the latest Opus model and moving up to Fable 5.2 only if tests at higher effort are not enough. For computer use the reasoning is the same, with one caveat: if the gap from Opus 5.5 is a few points on a public test, your own tests have to say whether it justifies double the price. The cost of an agent that watches the screen is also not just the model's: every screenshot is an image that counts as input, and long tasks generate many.

For model choice in general and a cost comparison, the reference guide is [Fable 5.2 vs Opus 5.5](https://giallostudio.it/en/blog/fable-5-2-vs-opus-5-5). For the big picture of what agents are and how they work, see [AI agents: what they are and how they work](https://giallostudio.it/en/blog/what-are-ai-agents).

## Where the first computer use pilot goes wrong

- **Using computer use where an API exists.** A direct integration is faster, cheaper and more stable than an agent that clicks. First check whether the program exports data or has an integration interface.
- **Trying it on the production environment.** A model error on real data is your error. Start on a copy.
- **Giving the agent the owner's credentials.** A dedicated, limited account makes every incident smaller and easier to reconstruct.
- **Measuring only the lucky case.** You need cases with exceptions: an empty field, an unexpected warning, a slow program. The real test is there.
- **Ignoring maintenance.** If the software changes version or a window moves, the flow can break. Decide who checks it and how often.
- **Leaving the old tool in place after moving to Fable 5.2.** With `computer_20251124` requests get the 400 error (except on Amazon Bedrock): it is the first thing to look for in the logs.

## How to get started in your company

The sensible path is short. Pick a repeated process that a person currently does on a program with no API. Check that an export or an API would not solve it better. Run twenty past cases in an isolated environment, with human approval on everything that writes. Then compare quality, time and cost with the current procedure, and only then decide whether to extend.

If you want to assess a real case, an office process to map before touching any tool, [Giallo Studio works on business process automation](https://giallostudio.it/en/business-process-automation) starting from exactly this kind of controlled trial. Before deciding, always re-check the [Anthropic official page](https://www.anthropic.com/claude-fable-and-mythos-5-2): versions, tools and prices change often, and for the wider context see also the guide to [AI Automation](https://giallostudio.it/en/blog/what-is-ai-automation).

## Frequently asked questions

### What is Claude computer use?

It is Claude's ability to work on a computer the way a person would: it looks at the screen and proposes mouse and keyboard actions on software that has no API. With Fable 5.2, released on 9 November 2026, it runs through the computer_toolset_20260801 toolset. It needs an environment (a virtual machine or a browser) that the developer sets up and controls.

### Why does the old computer_20251124 tool return a 400 error with Fable 5.2?

Because Fable 5.2 and Mythos 5.2 support computer use only through the new toolset. A request that declares computer_20251124 gets a 400 invalid_request_error on every platform except Amazon Bedrock, where the old tool still works. The fix is to declare computer_toolset_20260801.

### How good is Fable 5.2 at using a computer?

On OSWorld 2.1 Anthropic reports 89.3%, against 83.5% for Fable 5.1 and 87.2% for Opus 5.5. The figure refers to a partial version of the test on offline sets and is reported by the vendor, not replicated by third parties. It is not a success rate on your own software: measure it on your own cases.

### Can I let Claude use my office's management software that has no API?

Yes, as a controlled trial and not as an immediate replacement for people. Start with reading and preparation tasks (extracting data, filling in drafts) and have a person approve every action that changes or sends anything. If the software offers a reliable export or an API, those remain preferable.

### Is computer use safe for company data?

It depends on the environment you run it in. Screenshots and instructions are prompts: for Fable 5.2 Anthropic retains prompts and responses for at least 30 days, and zero data retention (ZDR) is not available unless authorized. You need an isolated environment, accounts with minimal permissions, no plaintext credentials and human control over actions that have consequences.

### Do I need computer use if I already have traditional automation (RPA)?

Not necessarily. If screens are stable and the steps are fixed, deterministic automation is more predictable and costs less per run. Computer use becomes interesting when the screen varies or when content has to be read and interpreted, accepting more variability.
