These are sixteen Codex projects, and every one of them is finished. OpenAI's Codex agent built eight briefs twice, once with GPT-6 Astra and once with GPT-6.1 Sol, from a single prompt each and with nobody editing the output. If you are looking for what people build with Codex, this is a controlled answer: the same briefs any other agent got, and the results side by side.
What did Codex build?
Eight kinds of project, each twice.
| Project | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|
| Marketing site | "Less busy. More in flow." | "Less busy. More you." |
| Analytics dashboard | "Good things are growing." | "Room to grow." |
| Physics simulation | Kinetic: "Order, in motion." | Kinetic: "Chaos, contained." |
| 3D shooter | Dead Signal | Dead Link |
| Phone finance app | "Hello, Jordan." | "Hey, Jordan." |
| One-lane MOBA | Emberfall | Aetherfall |
| Invented product site | PEEP, a bird whistle | AFTERHOURS, an off switch |
| Kanban board | Sidebar, green | Sidebar, orange |
The names alone tell you something: on most briefs the two models made strikingly similar choices.
How were these Codex projects made?
With the command published on our methodology page: Codex CLI in a fresh, empty folder, personal configuration switched off, the brief passed word for word, extra-high reasoning, a 150-minute cap. Each project took 7 to 25 minutes and ran as one agent turn of 10 to 33 shell commands. At list API prices, Astra's projects cost $2.12 to $7.27 each and Sol's $0.54 to $1.12.
What do Codex projects have in common?
Completeness and a house style. Every project includes every feature its brief listed, logged no runtime errors in our recordings, and stays compact, mostly one HTML file between 24 KB and 95 KB. The style is warm, tidy and conservative: serif display type, soft neutrals, a gentle motivational line in the interface. If you want a Codex project to stand out, ask for a direction explicitly.
How do Codex projects compare with Claude Code projects?
Faster and leaner against richer and slower. The same eight briefs built by Claude Code took 25 minutes to two and a half hours each and produced larger, more elaborate projects. See them on the Claude Opus 5.5 and Claude Fable 5.1 pages, or compare any brief side by side.
Four kanban boards, two built by Codex. Pick the one you would use, then reveal.
Take the blind test
Frequently asked questions
What kind of projects can Codex build?
Complete front-end projects from a single prompt: we have sixteen, including marketing sites, an analytics dashboard, a phone finance app, a kanban board, a physics simulation, a 3D shooter and a top-down MOBA. Each opens live on this site.
How long does a Codex project take?
Between 7 and 25 minutes per project in our runs, each in a single uninterrupted agent turn. The MOBA, the largest, took GPT-6 Astra about 25 minutes.
Were the Codex projects edited?
No. Each project is the files Codex left on disk after one attempt, published unchanged. The recordings use the same scripted input for every build on the site.
Which model did Codex use?
Two: GPT-6 Astra and GPT-6.1 Sol, both at extra-high reasoning effort. Each built all eight briefs, so you can compare them project by project.