
Claude Fable 5.2 Built an Operating System in Eight Days. Here’s What It Can Build for You
Your feed is full of demos powered by a thousand agents and $15,000 budgets, and you just want something that runs on your laptop. Here’s what Claude Fable 5.2 can actually do, where it beats Opus 5.5, and how to give it a real task in Unitool without overpaying.
A thousand-agent demo vs. your Tuesday night
Anthropic released Claude Fable 5.2 on November 9, and within a day X was flooded with clips that make you want to close your laptop and take up woodworking. A voxel world with real destruction physics, an Excel clone with pivot tables, a shooter that looks like it came from a big studio. The fine print is even more intimidating for a beginner. Fifty hours of model time, a team of agents, an estimated $15,000 at API prices.
The good news is that most of what makes Fable 5.2 useful doesn’t require a thousand agents, and the model is already available in Unitool. The less fun news is that it’s not a magic button, and Anthropic itself says you shouldn’t call it for every small job. Below we’ll cover what this model is, what the benchmarks really show, and how to hand it a task without paying flagship rates for an email to your coworker.
Fable and Mythos are one model behind two locks
Fable is the top tier of Claude, sitting above Opus. Mythos 5.2 launched alongside Fable 5.2, and they’re the same underlying model. The only difference is safeguards. Fable blocks high-risk cybersecurity and biology requests, which is why Anthropic can offer it to everyone. Mythos ships with some of those filters relaxed and is available only to vetted organizations such as research labs and security teams.
When Fable 5.2 declines a request, an earlier model answers instead, Opus 4.8 for cybersecurity and Opus 5 for biology. According to Anthropic, unnecessary blocks dropped by about 55% compared with Fable 5.1. For an everyday user the safeguards rarely get in the way, unless you’re trying to break into your neighbor’s server.
Anthropic also has numbers on speed and cost. Fable 5.2 generates output up to 30% faster than version 5.1 and costs less, at $8 per million input tokens and $40 per million output tokens on the API. A token is a small chunk of a word, the unit a model reads and writes in. There’s also a fast mode that speeds up responses by up to 2.5 times, but it costs twice as much and is offered in Claude Code and the developer API.
What Anthropic says Fable 5.2 can do
The announcement reads like a report from a small research institute. One Anthropic engineer gave the model an empty folder and asked it to design a prototype operating system. For eight days Fable 5.2 coordinated more than a thousand subagents (helpers it hands parts of the work to). The result boots on a laptop and has drivers for the disk, keyboard and network, permanent file storage, windows, a login screen and an installer.
From a one-paragraph prompt, the model designed three devices, a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. It wrote its own specs, picked parts from US distributors, routed four-layer circuit boards, modeled the enclosures and wrote the software. Each board was ready for manufacturing in under a day. The circuit board company Diode built them, and all three devices worked without any electrical modifications.
A team of 1,000 Fable 5.2 agents also formalized the proof of Fermat’s Last Theorem in 30 hours, translating it into a language a computer can check step by step. Anthropic’s first formalization took 11 days. Opus 5.5, running the same setup, got only 26% of the way there in 18 hours. And robots controlled by Fable 5.2 solved a puzzle cube and unscrewed a bottle cap, even though the model was never specifically trained for those tasks.
The benchmarks in plain English, awkward parts included
A benchmark is an exam for an AI model with answers known in advance. Every number below comes from Anthropic’s table, and the comparisons with GPT-6 Astra and Gemini 4 Carbon were run by Anthropic too. Keep that in mind until independent results come in.

FrontierCode 2.0 checks whether a model’s code changes would be accepted by the maintainers of real open-source projects. Fable 5.2 scores 49.1%, Opus 5.5 scores 45.5%, GPT-6 Astra 42.8%, and the previous Fable 5.1 just 32.4%. On FrontierSWE v2, another coding test, the new model hits 69.9% against 62.3% for Opus. On Humanity’s Last Exam, a set of extremely hard questions across many fields, it reaches 74.2% with tools versus 67.7%.
The biggest jumps show up on long tasks. On the Remote Labor Index, where the model completes real freelance-style jobs, Fable 5.2 scores 41.3% versus 21.3% for Opus 5.5, nearly double. The RoboDojo test, which has the model control a simulated two-arm robot, shows a similar gap at 64.9% versus 39.4%.
It doesn’t win everywhere. On AutomationBench (business workflows), Gemini 4 Carbon leads with 53.9% to Fable’s 51.4%. On the AA-Briefcase office-work ranking, Opus 5.5 has 1,800 Elo points and Fable 5.2 has 1,785. For emails, spreadsheets and reports you’ll barely notice a difference, and Opus is cheaper.
What Fable 5.2 costs in Unitool
Under the message box in the Claude Fable 5.2 chat you’ll see “Input 0.924 · output 4.621 credits per 1000 tokens.” For Claude Opus 5.5 it’s 0.605 and 3.025, so each Fable token costs roughly one and a half times more.
So why does Anthropic talk about savings? It comes down to reasoning effort. At High effort, Fable 5.2 solves FrontierCode tasks about as well as Opus 5.5 at Max effort (45.3% vs. 45.5%) while spending roughly half as much per task, because it needs fewer tokens to get there. In Unitool you set this in Chat Settings under Reasoning Depth, with five options from Low to Max. Thinking itself is always on for this model.

What people outside the lab have built
Take the X demos with a grain of salt. These are the creators’ own accounts, and we haven’t verified them. Matt Schumer, a serial entrepreneur in the field of artificial intelligence, a visionary, and an expert who has gained widespread recognition for creating innovative tools based on large language models, gave the new models a single task for a year—a voxel engine with physics—and, according to him, Fable 5.2 was the first to complete it with just one prompt. He then asked for a Harry Potter and Fortnite mashup and noticed the best results came when one agent ran the project from start to finish instead of handing tasks to helpers. Matthew Berman built an Excel clone in nearly 50 hours of model time. Ethan Mollick made a game where you play a ghost scaring off meddling kids, trailer included.
These stories have one thing in common. The creators showed up with an idea and a clear sense of what “done” looks like, and the model worked out the plan on its own.
A reading game in one evening with Unitool
Let’s pick something smaller than an operating system but still real. We want a browser game where a 6-year-old reads syllables and earns stars for correct answers. One of the creators on X built something similar for his own kid, and we’ll recreate the idea in Unitool.
Anthropic recommends starting with Opus 5.5 and moving up to Fable when the task gets too complex. That’s exactly what we’ll do.
Open Claude Opus 5.5 and ask it to write a spec. Prompt.
I want a browser game for a 6-year-old who is learning to read by syllables. Ask me 5 questions, then write a short spec: screens, rules, levels, and what counts as a finished version.
Click the model name at the top of the chat and choose Claude Fable 5.2. Open Chat Settings with the gear icon in the top right and set Reasoning Depth to High. In the menu under the message box (the sliders button), turn on Files and code so the model can run code, not just write it. Replies will take longer and use more tokens, as the setting itself warns.
Paste the spec. Prompt.
Build the game from this spec as a single HTML file with no external libraries. First outline your plan, then write the code, run it, check it against each point of the spec and list anything that didn’t work.
Save the code to a file ending in .html and open it in your browser. If something breaks, take a screenshot, attach it with the plus button (PNG or JPG up to 5 MB) and describe the problem in plain words.
In the same chat, turn on Unitool MCP and ask for a cover. Prompt. “Draw a cover for a kids’ reading game: a cheerful owl surrounded by stars, 1:1 format.” Before anything is generated, you’ll see the model, the settings and the price, and generation starts only after you confirm.
A cheaper model writes the spec, while Fable 5.2 handles the logic, code and testing. The cost of each request appears under the message box before you send it, so the expensive step never comes as a surprise.
MCP (Model Context Protocol) is a standard that lets a text model call other tools on its own. In Unitool it’s a toggle in the chat that gives the model access to image, video and audio models right from the conversation. The MCP settings include a Budget without confirmation field. At zero, every generation waits for you to click Run, and we suggest leaving it there until you get comfortable. If you work in Claude Code or Codex, connect Unitool MCP with your API key.

In practice, how to work with Fable 5.2
Give it a goal and a definition of done, not step-by-step instructions. “The game is finished when a child can clear three levels without help from an adult” works better than ten bullet points about buttons. Breaking down a fuzzy task is exactly what this model is good at.
Keep the project in one chat. Shumer’s observation lines up with Anthropic’s own pitch for Fable 5.2 as an orchestrator. When the whole history of the task is in front of the model, it loses the thread less often. Just remember that under Input Context, more history means a more accurate answer and a pricier request.
Match the reasoning depth to the job. Medium is enough for reviewing text or drafting a plan, High suits coding, and X-High or Max are for problems the other levels couldn’t crack.
Ask the model to check its own work and list what’s unfinished. In Anthropic’s testing, Fable and Mythos 5.2 were less likely to make up inputs or claim a task was complete when it wasn’t. Less likely isn’t never, so open the file and play the game yourself.
Why Fable, if Anthropic itself recommends Opus?
Fair question. Opus 5.5 is cheaper, and for most coding and document work it’s enough, as Anthropic says outright. We agree, which is why our walkthrough started there.
Fable pays off on long, messy tasks, like a program with several moving parts, research across piles of data, or a project other models abandon halfway. On those, every mistake means another round of fixes, and the higher price per token is offset by fewer attempts. Summaries, emails and social captions are still a fine job for Claude Sonnet 5.5 or Haiku 5.5. If you plan to use Claude every day, the Advanced plan at $16 a month gives you 1,000 tokens to spend across every model, so you can split them between Opus for drafts and Fable for the hard parts.
That idea you keep putting off
Most of us have an idea we’ve shelved for years because “you’d need a developer for that.” A calculator for your small business, a game for your kid, a spreadsheet analysis you never had the patience for. Fable 5.2 doesn’t promise a perfect result on the first try, but the barrier to entry is noticeably lower.
Pick one of those ideas, talk it through with Opus 5.5, then hand the hard part to Fable 5.2. The thousand agents can wait. One well-framed task is plenty to start with.
Get started with Claude Fable 5.2 in Unitool