Best game asset tools for Codex users

The character we used for every test

One brief, one character, four jobs. Everything below starts from the same robot explorer so the comparisons are about the tools rather than about four different prompts. The concept art itself is the first job: GPT Image 2 at $0.15 per run, asked for a full body in a neutral pose on a plain background — the shape a 3D model needs later.

1. Store art: the same character, a different frame

The job most indie projects hit first is store art, and the hard part is not making a nice picture — it is making the same character twice. Passing the concept image back as a reference did that. Two runs, two aspect ratios, $0.15 per run each, 65 and 75 seconds, and the robot is recognisably the same robot: same head shape, same orange panels, same face.

Two practical notes. Ask for the aspect ratio you actually publish at — 2:3 lands on the vertical library size exactly, while a 16:9 render needs a small crop for a wider capsule, and cropping a generated image is free local work. And keep text out of the image: models spell badly, and a title you add yourself in any layout tool stays editable when the price or the tagline changes.

2. 3D: the polygon budget is the whole decision

Both 3D services take one image and return a model, and on price they look four times apart. What actually separates them is what arrives. We ran the same concept image through both, on the cheapest configuration each offers.

That is the same character at 48 times the density on one side and a game-ready budget on the other, for a difference of about seventy-five cents. A 1.5-million-triangle mesh is not a bad model — it is a scan-grade one, useful if you are going to retopologise it yourself and wasteful if you are not. Both services accept a target polygon count and a face limit in the call, so the fix is to say the budget out loud in your request rather than to pick the cheaper service and then spend an evening decimating. Ask for triangles if you are heading straight into an engine and quads if the model goes through a modelling package first; the topology choice also decides the file format you get back.

3. Character voice: cheap enough to cast by ear

One line of robot dialogue, two speech models, same text. At $0.002 per 40 characters (text) of script, an 83-character line is $0.006 for 83 characters at $0.002 per 40 characters (text) — which means casting a character is a matter of listening to both rather than choosing on a spec sheet.

Both readings are usable and they are not the same performance: MiniMax lands slightly brisker, VoxCPM slightly softer. Neither needed a voice chosen or settings tuned. For a game with dozens of barks, the useful property is not which one wins but that a whole line list costs less than a coffee, so you can re-record after a script change instead of living with the old take.

How to run any of this from Codex

Ask in plain words. The Skill keeps an index of the catalog on disk, so before it runs anything the agent replies with the services that fit, each with its price, and picking one is the approval. Nothing is charged until you pick, and if the wallet cannot cover the run the agent stops before anything is spent and hands you a top-up link. The image work goes through the image generation capability, the models through 3D generation, and the dialogue through voice generation — one balance covers all three, and every run lands as a file in your project.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.codex/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.codex/skills/actionway"

You need Node.js 20 or newer and an Actionway account. Prefer a plugin or a Remote MCP server to the command line? Both end at the same catalog and the same wallet, and the get started page lists the steps per client side by side. Claude Code users install the same Skill into ~/.claude/skills/actionway instead — the image editing guide walks through that route with a product photo.