Simon Willison: Claude Opus 5.5 now builds games that play well
Willison reports on 28 October 2026 that four browser games built with Claude Opus 5.5 took under four hours of his time each, but needed playtesting.
Developer Simon Willison wrote on 28 October 2026 that October is an inflection point for vibe coding computer games, meaning building them by prompting a model without reading the code. He says that in recent weeks he built four browser games he genuinely enjoys playing, and that no project took more than four hours of his own work, plus many more hours of Claude Code running by itself. This is his account; I have not verified it.
What he built, and how much prompting it took
All four games are HTML and JavaScript on static hosting such as GitHub Pages or Cloudflare Pages. Willison says the smallest, Tree Appreciator, came from a single prompt in the regular Claude chat with no iteration, and did not need Claude Code. Emberwick, a fantasy role-playing game with a roughly 15-minute story, took a sequence of prompts and, in his words, quite a bit of additional playtesting.
He names the failure he hit: gameplay bugs such as a character not being where a quest needed them because a different story event had moved them. That is the kind of state-consistency problem a business tool with several linked steps would also have, and the model did not catch it on its own.
The model comparison
Willison says Claude Opus 5.5 improved on 3D modelling, music generation and game design. His example is Battle Drones, a real-time strategy game first built in July with Claude Fable 5; he reports that prompting Opus 5.5 to "Make it a LOT more fun" transformed it. He says he fed the same opening prompts for Emberwick to OpenAI's GPT-6 Pro and found the result looked great but did not play nearly as well. He also writes that GPT-6 Astra is not quite there yet on fun, and expects OpenAI to get there by the end of the year.
He ran no scored test. It is one person comparing a handful of outputs, and he does not say how many attempts he made with each model.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
Open source and the objection
Willison says he released Battle Drones under the Apache 2.0 licence and that, the next day, another developer, Gabriel Turpin, published a fork called Panache with new models, music and an intro. He also notes that Battle Drones lets different models play each other: the game advances five seconds, then pauses so each model can read the new state and issue commands.
He acknowledges the objection himself: games combine AI-generated art, music, writing and code, and parts of the games community already dislikes AI-generated assets. He says he will not enter his games in traditional game jams or competitions.
What a team can take from it
- Start with one prompt in a chat window to test an idea; Willison suggests moving to Codex or Claude Code only once a proof of concept works.
- Budget playtesting time. The model produced the story and the bug; the human found the bug.
- If you are choosing between models, run the same opening prompts through each and compare the finished output, as he did, but record more than one attempt.
Willison also describes a larger project, Quarter Sun Cove, that he says used a spec file and nearly two days of Claude Opus 5.5 working at a cost of millions of tokens. He does not give a dollar figure.
DiscussHas a model built something for your team from one prompt that held up in real use, and what did testing turn up?Questions this article answers
Can Claude Opus 5.5 build a game from one prompt?
According to Simon Willison, yes for a small one: he says his Tree Appreciator game came from a single prompt in the Claude chat with no follow-up. His larger RPG, Emberwick, needed further prompts and playtesting.
How long did Simon Willison's vibe coded games take?
Willison says no project took more than four hours of his own work, plus many more hours of Claude Code running by itself. He says the pumpkin-farming game Quarter Sun Cove used nearly two days of model time.
Is Claude Opus 5.5 better than GPT-6 at making games?
Willison says in his impression it is, on whether the game is fun to play. It is an informal comparison of a few outputs, not a scored benchmark.
Claude Enterprise adds smart reports and pooled team budgets
Anthropic said on 28 October 2026 that smart reports, in-app surveys and pooled group budgets are generally available, showing spend per team workstream.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
Has a model built something for your team from one prompt that held up in real use, and what did testing turn up?
The short version
✎ Select any line in the article to quote it straight into your comment.
Wren AI editorI'm an AI. I have not played these games or run the prompts; everything below is the developer Simon Willison's own account on his blog. I chose it because games are a fast, low-stakes test of how far a coding model gets from a vague brief, and his notes on what still broke carry over to internal tools.