OpenAI and ModRetro connect Codex to a $200 retro console for making your own games

Asking AI to write a game is easy. Getting it to the point where someone else can launch it and figure it out without your explanations is harder. That stretch of development is what makes the collaboration between OpenAI and ModRetro interesting: the proposed workflow takes an idea through Codex to a game you can test and then transfer to a cartridge. For developers, it offers a chance to examine how smoothly the process moves from generation to testing the result.

The ModRetro Chromatic: Codex Edition costs $200 and comes with a programmable cartridge. It is a version of ModRetro's console released as part of its collaboration with OpenAI. The proposed workflow uses Codex, the ModRetro Chromatic plugin and the GB Studio editor to work on scenes, graphics and game logic. Details are available on the OpenAI project page.

From prompt to playable game

Imagine a task: make a short platformer in which a character collects items and reaches an exit. This is a hypothetical example for testing the process. Keep the initial prompt to one scene, one action and a clear win condition. That makes it easier to see what worked and where the agent misunderstood the idea.

The substantive work begins after the first build. Does the character jump too far? Does an item block a passage? Is there a long wait after losing? Each observation can become a separate task for the next revision. The tool's value to a team will depend on how predictably it makes these changes while preserving mechanics that already work.

GB Studio lets you run a built game from the editor and start testing from a selected scene. Its export options let you create a ROM for a compatible emulator or a web version to publish on a server. These build options are described in the GB Studio documentation. For a small prototype, this is a convenient handoff point: you can give a tester a runnable build and a specific assignment.

OpenAI also describes testing a game on Chromatic from a computer and running it independently from a cartridge. Before using the cartridge, you need to update the firmware and enable developer mode. In other words, the usual technical steps still stand between a successful prompt and a game running on the device, and they need to be completed and checked.

What you can try without buying the console

You can start with OpenAI's gallery. It features Game Boy Color games created with Codex, including FlapGPT, Codex Land and Hollow Descent. They are playable in a browser, and ROMs can be downloaded for compatible hardware.

The gallery helps you assess the scope of the examples and how the games feel to play. But a finished result tells you nothing about the number of failed attempts, the creator's time or the amount of manual refinement. The cited sources provide no measurements that would support promises of a specific development speedup or task success rate.

For your own assessment, it is therefore more useful to repeat a small experiment. Define the mechanic in advance, build a first version and ask someone else to play through it. Record the problems they encounter, how many fixes are needed and how long it takes to test each build. That data will be more useful than the impression left by a polished demo.

Why developers and product managers should care

For a developer, a project like this can be an exercise in defining tasks for an agent. A short game makes it easy to formulate testable requirements: an item can be picked up, a door opens, losing restarts the scene. Each requirement can be checked through an action, allowing you to determine whether Codex fixed a specific bug.

Product managers should look at the whole cycle: how quickly a version ready to show someone emerged, and how costly it was to change after feedback. For founders, the same experiment can help assess an agent's usefulness for small interactive prototypes. Applying those conclusions to complex commercial games without separate testing would be premature.

A sensible first step is to choose one mechanic and develop it into a browser version that feels good to use. You can then decide whether your project needs a physical cartridge. The collaboration is interesting precisely because it lets you go all the way to a runnable product and see where human intervention is still needed.