---
title: "How Did Fable 5.2 Coordinate 1,000 Agents to Build an OS?"
url: "https://goremy.ai/blog/articles/fable-5-2-1000-agents-build-operating-system"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-11T05:20:10.144Z"
updated: "2026-11-11T05:20:17.096Z"
topic: AI Agents & Automation
description: "Anthropic says Fable 5.2 ran 1,000+ sub-agents for 8 days to build a bootable OS from an empty folder. Here's what that shows and what it means for you."
---

# How Did Fable 5.2 Coordinate 1,000 Agents to Build an OS?

> Anthropic says Fable 5.2 ran 1,000+ sub-agents for 8 days to build a bootable OS from an empty folder. Here's what that shows and what it means for you.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-11

## How did Fable 5.2 coordinate over 1,000 agents to build an operating system?

According to Anthropic, one of its engineers gave Claude Fable 5.2 an empty folder and the task of designing a prototype operating system. The model coordinated more than 1,000 sub-agents, which worked around the clock for eight days writing, testing and reviewing code. The result boots on a laptop. It has its own core, drivers for the disk, keyboard and network, permanent file storage, windows, a login screen, an installer and built-in Claude functionality. Anthropic has not published the coordination mechanics in the material we reviewed.

## TL;DR

- Anthropic reports that Fable 5.2 ran **more than 1,000 sub-agents for eight days** to produce a bootable prototype operating system from an empty folder.
- A second demonstration, also from Anthropic, had a similar-sized group of agents turn the proof of Fermat's Last Theorem into **code a computer can check, in about 30 hours**.
- The point of both demos is **long-horizon work**: tasks too big for one chat session, where planning, testing and review matter as much as writing.
- These are **Anthropic's own showcase examples**, not independent tests, and the OS is a prototype.
- Anthropic itself recommends **Opus 5.5 for most coding and knowledge work**, and positions Fable 5.2 for the most ambitious, open-ended technical problems.
- You do not need 1,000 agents to benefit. The useful idea is **splitting big work into pieces, with checks built in**.

## What exactly did the operating system demo include?

The list Anthropic gives is specific, which is what makes it interesting. Starting from nothing, the finished prototype has:

- A core (the part of an operating system that runs the machine).
- Drivers for the disk, keyboard and network. Drivers are the small programs that let software talk to hardware.
- Permanent file storage, so files survive a restart.
- Windows, meaning a graphical interface.
- A login screen.
- An installer.
- Built-in Claude functionality.

That is a lot of separate pieces that have to agree with each other. A disk driver that works alone is useless if the file system expects something different. Building an OS is a good stress test because errors in one area show up as crashes in another.

It still has limits. Anthropic calls it a prototype, and nothing in the sources compares it to an operating system you would install on a work laptop. One reviewer who covered the launch also pointed out that these are Anthropic's cherry-picked examples. Treat them as a demonstration of what is possible, not a guarantee of what you will get on your own task.

## What was the Fermat's Last Theorem effort?

Fermat's Last Theorem is a famous mathematics problem, and its published proof is long and complex. Anthropic says a similarly large group of Fable 5.2 agents turned that proof into code a computer can check, and did it in 30 hours.

This is called formalization. A human-written proof relies on readers to follow the logic. A formal proof is written so a program can verify every step, with no gaps and no "this is obvious." It is slow, painstaking work, which is why it makes a good long-horizon test.

Notice the shared feature with the OS demo: both have a built-in way to tell right from wrong. An OS either boots or it does not. A formal proof either passes the checker or it does not. That kind of clear pass/fail signal matters a great deal for agents, as the next sections explain.

## How does coordinating hundreds of agents actually work?

The sources do not describe Anthropic's exact setup, so here is the general pattern rather than a claim about their system.

Multi-agent work usually has three roles:

1. **A planner or orchestrator** breaks the goal into smaller jobs and hands them out.
2. **Worker agents** each take one job, such as writing a network driver or a login screen.
3. **Reviewers and testers** check what the workers produce and send problems back.

Anthropic's description, "writing, testing, and reviewing code," matches this shape. The hard part is not having many agents. It is keeping them consistent over eight days, so that the work of agent 400 still fits with what agent 12 did on day one.

Anthropic also suggests using Fable 5.2 as an advisor or as an orchestrator. In that setup, it delegates simpler tasks to smaller models (Haiku, Sonnet or Opus) rather than doing everything itself. That is a sensible way to think about cost: use the expensive model for planning and judgment, and cheaper ones for routine work.

## Why does long-horizon orchestration matter?

Most people use AI in short bursts: a question, an answer, a revision. Long-horizon work is different. The model has to stay on task for hours or days, track what is finished, recover from its own mistakes and keep the pieces compatible.

One reviewer described Fable 5.2 as especially strong on the "big hairy problems" that span a long time and involve lots of complexity. Anthropic's own benchmark figures, as reported, point the same way. On an agentic scientific research benchmark, the reported scores were 58.7 for Opus 5.5 and 74.6 for Fable 5.2. On everyday coding the gains were smaller: one frontier coding benchmark went from 45.5% to 49.1%.

The pattern is that the bigger and messier the job, the more a stronger model helps. For a quick email or a spreadsheet formula, you will not notice much.

## Did Anthropic hold this model back?

Yes, in specific areas. Anthropic says that during training it took steps to limit advances in research capabilities that could most directly accelerate AI development. Two examples named in coverage are accelerator kernel optimization and improving large language model architectures. In those areas, Fable 5.2 performs about the same as the cheaper Opus 5.5.

Anthropic says this is a training decision, not something applied while you chat with the model. For most workplace tasks you are unlikely to hit it. It does show that labs are now choosing what to include, which is new.

## How much does it cost, and who can use it?

Fable 5.2 is available on paid Claude plans and through the API. It is not available on the free plan. API pricing as reported:

| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Fable 5.2 | $8 | $40 |
| Fable 5.1 | $10 | $50 |

That is 20% less than the last Fable and, per one reviewer, double the price of Opus 5.5. Artificial Analysis, an independent benchmarking site, put the average cost of a task at about $26.54 on Fable 5.2 at its maximum reasoning setting. The same review noted diminishing returns there: on one coding benchmark, the "extra high" setting scored better than "max" while costing less. Running a model for eight days with 1,000 agents is therefore a big spend, and not something a typical office user would do.

There is also Mythos 5.2, a version with fewer cyber and biology safeguards, available only to verified organizations.

## What does this mean for you?

**If you manage projects or teams:** the lesson is how the work was structured, not the agent count. Big goals were split into pieces, and every piece had testing and review attached. You can borrow that. When you hand a large task to an AI tool, break it into stages and define what "done and correct" looks like for each.

**If you write code or work near software:** start with Opus 5.5, as Anthropic recommends, and move to Fable 5.2 when the problem is large, open-ended and stuck. Try a medium or high reasoning setting before max.

**If you are worried about your job:** the demos show AI handling big, structured, checkable technical projects. They say little about work that depends on judgment, relationships or unclear goals. The tasks that look most exposed are the ones with a clear pass/fail test.

**What to do this week:**

1. Pick one multi-step task you do regularly, such as a monthly report.
2. Write down the stages and how you would check each one.
3. Try an AI tool on one stage at a time, with that check in place.

## Can you build tools like this without 1,000 agents?

Most people's version of this is smaller: a request form that routes work, a tracker that shows what is waiting for review, a dashboard on top of a messy spreadsheet. [Remy](https://goremy.ai) builds and deploys that kind of full-stack app from a plain-English description. Remy works from a written spec, then builds the app, tests it and puts it on a live URL. Apps can also run task agents that research or generate content and write results back to the app's own data. If you are juggling spreadsheets for a review process, describing the workflow and getting a working internal tool may be less effort than buying another subscription.

## Frequently Asked Questions

### Is the operating system Fable 5.2 built ready to replace Windows or macOS?

No. Anthropic describes it as a prototype. It boots on a laptop and includes the core pieces of an OS, but nothing in the sources suggests it is a finished product for everyday use. The demo shows capability, not readiness.

### Can I run 1,000 agents myself?

The sources do not explain how Anthropic set this up, so we cannot say how to reproduce it. Fable 5.2 is available through the API on paid plans, but a run of this size would be costly, given that a single maximum-effort task averaged about $26.54 in one independent analysis.

### Should I use Fable 5.2 instead of Opus 5.5?

For most people, no. Anthropic recommends starting with Opus 5.5 for most coding and knowledge tasks and moving to Fable 5.2 for the hardest, most open-ended technical problems. Fable 5.2 costs about double per token.

### What is formalizing a proof?

It means rewriting a mathematical proof in a precise language so a computer can verify every step. Anthropic says Fable 5.2 agents did this for the proof of Fermat's Last Theorem in about 30 hours.

### Are these results independently verified?

The two big examples come from Anthropic's own materials. One reviewer explicitly called them cherry-picked. Independent benchmark sites such as Artificial Analysis have tested the model more broadly, but the OS and proof projects are Anthropic's own demonstrations.