A single developer used Claude Fable 5, an unreleased AI model from the makers of Claude, to rewrite an entire piece of software from scratch in just 11 days. Over a million lines of code, mostly written by the AI, for a total bill of around 165 000 dollars in API fees.
The Gist
- Bun, a modern developer tool for building websites and apps, was fully rewritten from one programming language to another using Claude Fable 5.
- The rewrite took 11 days, ran up to 64 Claude instances in parallel, and cost roughly 165 000 dollars in API fees.
- The new version fixes 128 bugs and runs 2 to 5 percent faster.
- The story shows what happens when a beginner-friendly idea (AI helping with code) gets pushed to industrial scale.
Have ChatGPT Recap This Article
ChatGPTMeet Bun and Fable 5, the tool and the AI behind it
Bun is a piece of software that developers use to build modern websites and apps. If you have ever used a service like Notion, Vercel or a startup app that feels snappy, chances are Bun or one of its cousins runs somewhere in the background. It is quiet infrastructure, but it matters.
Claude Fable 5 is the codename Anthropic uses for one of its most advanced AI models, not yet released to the public. It is the same family as the Claude you may already know, but tuned specifically to handle very long tasks, including writing and reviewing large amounts of code without dropping the thread.
The story broke thanks to Jarred Sumner, the developer behind Bun. He shared his own numbers publicly. Between May 3 and May 14, a small team used a pre-release version of Fable 5 to translate Bun’s entire codebase from one programming language, Zig, to another, Rust. The final diff includes 535 496 lines of the old code translated and over one million new lines added.
If you want a broader map of why Anthropic is suddenly the loudest name in AI right now, our beginner-friendly explainer on Anthropic quietly making more money than the maker of ChatGPT is a five-minute read. Fable 5 is the kind of muscle that story runs on.
How Claude wrote a million lines of code in 11 days
The trick was not one Claude answering one question at a time. The team spun up as many as 64 Claude instances working in parallel, each one handling a different piece of the codebase, all coordinated by one human engineer keeping an eye on the whole. Think of a warehouse where 64 clerks pack in parallel while one manager walks the aisle.
The final tally is impressive on every axis. The AI produced 6 502 commits, which is the coder’s word for saved checkpoints. It ran 1 386 826 automated tests, checking that every piece still worked as expected. Zero tests were skipped. That is not a demo, that is a shipped product.
The final result, called Bun v1.4.0, fixes 128 bugs the old version had and runs 2 to 5 percent faster. The binary size, which is the space the software takes on your machine, drops by roughly 20 percent. In other words, the AI did not just clone the tool, it delivered a cleaner one.
On the flip side, this is not the story of one AI replacing a human team overnight. It is one human engineer, deeply familiar with the code, using AI as a workforce that never sleeps. The judgment stays human, the typing gets outsourced.
Keep learning on AI Noobies:
- Muse Spark: Meta’s New AI Model That Uses Your Apps
- Claude Reflect Shows How You Actually Use AI
- 5 Common ChatGPT Mistakes Beginners Make Daily
What “parallel instances” and “API cost” actually mean
Two terms decide whether this story is an expensive stunt or a repeatable move. The first is running instances in parallel. You are not paying one AI, you are paying 64 at the same time. When each one thinks for a few hours across a big task, the meter runs fast.
The second is the cost of the API. AI companies bill by tokens, which are chunks of text sent and received. Fable 5 read the equivalent of 5,9 billion tokens as input and wrote 690 million as output during the 11 days. The final invoice came out at roughly 165 000 dollars. If you want a friendly primer on what a token actually is, our beginner guide to tokens in AI covers it.
Set against a human team doing the same work, the money reads very small. A rewrite of that size, done by three or four senior developers over six months, would cost several times more in salaries. That is the reason this story is being read closely inside every software company right now.
The nuance here is quality control. Anthropic’s tool is designed to catch its own mistakes with adversarial reviews, which is a fancy way of saying one Claude checks the work of another Claude before the code is merged. It sounds fragile, but on this project it caught real bugs before they made it into the final tool.
Where you might see this shift in your own tools
You do not need to touch code to feel the impact of what happened here. When the tools that build your apps get faster, cleaner and cheaper to ship, the updates you see on your phone speed up. New features arrive sooner. Weird bugs get fixed in days instead of months. That is the direct beginner-side consequence.
Zooming out, the wider public debate is starting to shift too. The question is no longer whether AI can write code, it is how much of the industrial pipeline can be handed over. Expect newspapers to pick up the same 11 day, 1 million lines story over the next weeks, usually framed as either a productivity revolution or a warning about developer jobs.
If you want to see the same shift happen with less code and more clicking, our beginner primer on what AI agents really do in the Copilot AutoPilot family is a good side-by-side read. Same idea, different surface.
A good first personal test is to try one of the free Claude chats. Give it a small automation task, like drafting an email to your landlord, then a rental notice, then a short reply. Watch how it holds the thread across the three. That thread-holding is the same skill Fable 5 used at scale on Bun. Once you feel it in a small task, the 11 day rewrite story stops sounding like science fiction and starts sounding like a size change.
Stay tuned on AI Noobies.
About the Author
