Project story / Ship a Game · 19 Jul 2026
Can an AI model research, build and ship a fun iPhone game?
Every new frontier model arrives with a fresh set of benchmark scores. Soon afterwards, YouTube fills with people asking it to build a web game. The results are often impressive for a few minutes, but they leave out most of the difficult questions. Did the model research what makes the genre work? Is the game still interesting after the novelty wears off? Can it handle assets, testing, store requirements and the long tail of getting a real product ready to submit?
I wanted to test the whole journey. Ship a Game gives a model one challenge: research, design, build and prepare a paid iOS game for App Store Connect. The human only handles the account actions Apple requires from the account holder and the final submission tap.
The intervention log is the score
A coding benchmark can hide a surprising amount of help. Someone fixes the environment, clarifies the prompt, chooses the engine or rescues a broken build, and the final screenshot still looks like a one-shot success. In this benchmark every intervention is part of the result.
I record nudges, fixes and rescues. A model that makes a good game after several targeted corrections has demonstrated something useful, but it has not demonstrated full autonomy. The log makes that distinction legible. It also makes failure interesting: the point is to see where the current boundary really is, not to manufacture a win for a preferred model.
Each model runs in its own first-party coding environment on the same provisioned Mac, with the same broad capability surface and the same challenge. It can choose its engine and its approach. The benchmark therefore tests research, planning, taste, tool use, debugging and delivery together.
The surprising answer so far
The models can build the games. Ringbloom, built through Codex, and Brinkball, started in Claude Code, are both live on the App Store. Brinkball reached its first submission with two nudges and no fixes or rescues; its later improvement round also exposed an important model-fallback complication that the full report documents. That is a much stronger—and messier—result than generating a playable web prototype in a browser tab.
The more interesting question is whether the world finds the games fun. I do, but I am both the person running the experiment and someone unusually interested in the process. Store approval, purchases, play time and retention are harder tests than my own enthusiasm. Those results need to remain separate from the autonomy score: a model can ship efficiently and still make a mediocre game, or need help while making something people love.
Shipping is one result. Making something strangers choose to keep playing is another.
Why do the models keep making puzzle games?
There is already a pattern worth pushing on. Turn-based and puzzle-like designs are a rational choice for an autonomous model: the rules are bounded, the state is testable and the game does not depend on frame-perfect feel. That makes them excellent evidence of software delivery and a possible escape hatch from harder kinds of game design.
A benchmark should not punish a model for making a sensible decision, but it should eventually test whether that decision has become a rut. Future prompts may need to constrain the genre or explicitly reward forms of play that rely on motion, timing or a less easily simulated kind of fun.
What I am really measuring
This started as curiosity about model capability. It has become a compact test of end-to-end product work. Writing code is only one part. The model has to decide what to make, recognise weak output, recover from tools that do not behave as expected and finish a process full of unglamorous details.
The first runs suggest the frontier has moved well beyond toy prototypes. They do not prove that an AI can independently make a successful game. That gap — between technically shipped and genuinely wanted — is exactly why I want to keep running the experiment.
Explore the live benchmark ↗ · Read the Brinkball report ↗ · Read the Ringbloom report ↗