Building an application with ChatGPT: what you get and what is missing

Building an application with ChatGPT works up to an identifiable point. What the conversation provides, what OpenAI's coding agent adds, what it costs, and the three things it will not do for you.

Building an application with ChatGPT produces working code, provided you take it outside the conversation. OpenAI's coding agent, Codex, has no separate price: it is paid through ChatGPT subscriptions (learn.chatgpt.com, pricing recorded September 2, 2026: Free, Go at $8, Plus at $20, Pro from $100). It assumes a terminal.

What conversation alone gives you

In the chat window, ChatGPT writes code you copy elsewhere. For a spreadsheet formula, small script, or page showing an idea to a partner, the service is real and immediate. The wall comes when the project has more than one file: the conversation does not see your folder, execute locally, or know yesterday's version worked. You become the thread linking the pieces, and that stitching is the real cost.

Codex, the version touching your files

Codex exists to remove that stitching. It works locally in an existing code folder, runs tests, fixes, and repeats; another version executes on OpenAI infrastructure and returns a change proposal to review. Two details matter to whoever pays. Your text and file content go to OpenAI in both cases, including when the agent runs on your Mac. ChatGPT and Codex usage share a five-hour window, so coding consumes your conversational quota.

The September 1, 2026 addition in official release notes is Vim-mode search in the draft editor. Such a line reveals the audience: developers settled in their habits, far from a business owner wanting management software.

Quota sharing works both ways. A day of agent corrections leaves you without conversation that evening, and the limit arrives mid-task without warning of what remains. For professional use, that alone can prompt upgrading, making the displayed $20 an entry price rather than the year's actual price.

The three things it does not do

Codex takes instructions rather than an approved specification: nothing requires it to write what the product must do before building, and nothing preserves that decision. It runs existing tests without requiring they precede code. It asks permission per action and shows changes to review, different from a gate before a document accepted or rejected before continuation. Those gaps describe its audience: someone already knowing what they want and able to read what returns.

For someone who will not code

If comfortable with a terminal, Codex offers the market's best value: one bill covers conversations and coding agent, no extra layer is needed, and approval gates would be pure friction. We say so plainly, as in four routes to an application without knowing how to code. If the terminal is a wall, it remains one: opening a black window, installing a package, understanding a version folder, reviewing changes line by line. Maestro controls Codex and three other assistants without you opening any of that, adding a brief, specification, step breakdown, each approved by you, and tests written before code. The layer costs an extra dependency and tool, while choosing Codex or Claude Code becomes a setting rather than a commitment.

Test your idea in an hour without installing anything

Open a conversation and describe your need in five sentences including special cases. Ask for a specification rather than code: screens, data, who sees what, what happens when an order is cancelled. Read it in ten minutes. Repeat a week later with the same five sentences and compare: differences reveal what the conversation invents when you have not specified it. If it surprises you, the description lacked precision, and you learned more than a mockup would teach. This is the boundary between prototype and product, crossed through text rather than screens.