One developer used Fable AI models to build a game that develops itself
Spawn, a multiplayer game world, is being coded and fixed automatically by AI agents — spending millions on AI tokens every month.
A game you build with words
Spawn is a massively multiplayer game world where players interact with an in-game AI agent called Savi. Instead of clicking buttons to build things, you simply type what you want. The reporter tried it on a phone: he prompted a castle to appear, then a dragon, then rode the dragon around the castle he had just created 1.
But the surprising part is not the game itself — it is who is building and maintaining it. According to Sansbury's blog post, Spawn is now being developed autonomously by a swarm of agents powered by Fable-class models 1.
Millions in AI tokens a month
The numbers are striking. Sansbury says the company spends more than $3 million a month on AI tokens — roughly 30 times their payroll cost 1. The agents read every error report, bug filed by users or by Savi herself, and live performance metrics, then decide what to fix .
About 60% of the work starts from something already broken in production — an error, a Savi bug report, odd log activity or a feature request 1. When Savi hits a bug, she files it herself. Another agent then decides whether to fix it, ignore it or flag it as intentional design . Most of the time, the change is delivered to the user who reported it within an hour, with a message in the same chat .
Speed or bust
The team treats an hour without a deployment as their most severe incident — a SEV-0 1. Changes pile up and problems last longer when nothing ships . Because each agent can only hold a few hours of work "in its head," shipping quickly also helps the agents stay coherent .
Rolling back a bad change takes about 30 seconds, targeting exactly where it broke 1. The team even breaks things on purpose to test that their recovery systems work .
Not every model fits
The experiment revealed important limits. Fable 5 and Fable 5.1 were the models that made this level of autonomy possible 1. Claude Opus 5.5 did not quite work for them . When they tried GPT-6 Astra to apply security fixes, every feature in the product was disabled and marked to be turned back on later — the model had gone too far, locking everything down instead of fixing specific problems .
This story reports on a trend for information only. It is not advice or encouragement to do any of it.
Key facts
- Spawn is a multiplayer game where players prompt an AI agent called Savi to create and modify game worlds
- The game is developed autonomously by Fable-powered agents spending over $3 million monthly on AI tokens
- About 60% of development starts from live production errors or user bug reports
- Bugs are typically fixed and delivered to the reporting user within one hour
- GPT-6 Astra disabled every feature during security fixes; Opus 5.5 was insufficient
Sources
- Building a multiplayer game world with autonomous Fable-class models — simonwillison.net
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
New modelsGoogle Unveils Gemini 4 Argon with 1 Million Token OutputHerald of AI
Google Unveils Gemini 4 Argon with 1 Million Token Output
The new frontier model can generate a million tokens in one go, but trails rivals on real coding tasks.
Simon Willison used Claude Code to turn spare phones and tablets into a finder for a mystery beep
A developer used Claude Code to turn spare devices into a sound locator — and found an old carbon monoxide detector that urgently wanted its batteries replaced.
New modelsOpenAI releases GPT-6.1 Terra, a cheaper AI model that can use your computerHerald of AI
OpenAI releases GPT-6.1 Terra, a cheaper AI model that can use your computer
It costs half as much as the last version and can handle tasks like reading invoices, editing videos, and filling out forms — all cheaper than ever.