If ChatGPT Sites actually builds or edits a site, the work counts toward Codex usage. Discussing a game's plot in an ordinary chat is a different activity. The distinction is what gets executed, not which text box looks more convenient.
Imagine asking an AI to build a browser RPG. You try @Sites in a regular chat and see progress. You try Work and see progress there too. Suddenly you feel like you've discovered a secret game exploit: two entrances must mean two free allowances, right?
The final boss of this theory is the usage meter.
What OpenAI explicitly says
OpenAI's Sites help page states that building or editing a Site counts toward Codex usage. Running and hosting a site, including traffic and storage, is handled separately.
Three situations should not be mixed up:
- Brainstorming characters, dialogue or gameplay in ordinary Chat is conversation, not automatically a Site build.
- Generating or modifying a real Site through @Sites is Site-building activity and uses the relevant Codex allowance.
- Creating or editing a Site through Work or Codex is also Site-building activity.
This is about the operation. Moving the same task into a different-looking chat window does not establish a second allowance.
Can ordinary Chat launch Sites?
The official creation workflow starts with Work on the web, or Work or Codex in the desktop app. It tells users to request a website or mention @Sites. If a normal Chat interface seems to be doing something, that observation alone does not prove the complete Sites workflow ran.
Check for an actual, editable Site preview or generated Site, rather than treating “I'll build it” as proof. And running two attempts does not mean their projects automatically merge into one magnificent two-headed RPG.
Unlimited Sites does not mean unlimited building
For Plus and Pro, OpenAI lists no fixed cap on the number of Sites. Usage and storage limits still apply. The number of projects you can keep is not the amount of work you can request today.
Work and Codex share an allowance, and consumption depends on the model, task and settings. This is not a guaranteed flat fee per website. Look in Settings → Usage for available usage and reset information.
There is also a timing distinction: as of November 2026, OpenAI has announced that eligible ChatGPT and Codex allowances will become shared starting in December. Do not describe that planned change as if it had already happened in November; it is separate from the existing rule that Site creation and editing use Codex usage.
A practical plan for a ridiculous RPG
Suppose your game has a fantasy town where the NPCs gradually start speaking English while the protagonist stubbornly stays in Japanese. Develop the scenes, branching choices and battle rules in conversation. Then submit a consolidated build request to Sites. Inspect the preview, report concrete bugs and iterate.
This will not guarantee a particular amount of savings, but it can avoid regenerating the same work in competing sessions. If you compare entry points, use small, similar test tasks and inspect the actual usage indicators. Different models or requirements make raw comparisons misleading.
The punchline
Two doors to the workshop are not two wallets. Conversation is conversation; actual Site construction or editing counts toward Codex usage. Verify the artifact in the preview and the allowance in Usage. You've unlocked a secret passage—straight back to the same meter.
