Codex: the coding agent included in your ChatGPT subscription
- to get started
- a ChatGPT plan (even Free, with tight limits) or an API key · CLI, web, IDE or iOS
- when to pick it
- You already have a ChatGPT subscription and want a coding agent without adding another fixed cost.
- official site
- developers.openai.com
pricing
| plan | checked on | |
|---|---|---|
| ChatGPT Plus | $20/month, includes Codex on web, CLI, editor and iOS | 19/07/2026 |
| ChatGPT Pro | $100 or $200/month, with 5x or 20x Codex usage compared to Plus | 19/07/2026 |
| API | pay per token at standard OpenAI rates, no hourly windows | 19/07/2026 |
Codex is OpenAI’s coding agent, and its most relevant commercial trait is that it lives inside the ChatGPT subscription: if you already have a plan (from Free with tight limits up to Pro, Business and Enterprise), Codex comes with it, on the web, in the terminal, as an editor extension and even on iPhone. The CLI software is free and open source; what you consume is your plan’s quota.
This site’s radar crosses paths with it constantly (it is one of the most used agents in the world, and the numbers behind its growth are a story we covered): this sheet is the service piece that was missing.
How billing works
Codex usage on ChatGPT plans draws from the same five-hour-window pool as the rest of the plan: a heavy CLI session and a heavy web session compete for the same quota. Plus ($20/month) covers focused sessions; the Pro plans ($100 or $200/month) give 5 or 20 times that usage. Alternatively, an API key pays per token at standard rates, without windows. Figures checked on 19/07/2026 on the Codex pricing page; as with the other agents, the precise window limits change often and should be re-read there.
Where it earns its keep
Integration with the OpenAI ecosystem is the strong point: same account, same models, and moving between asking in chat and delegating to the agent feels natural. For teams on ChatGPT Business, the Codex quota arrives with the seats already paid for. The reasoning for choosing it is often economic before it is technical: given a subscription you already hold, trying it costs zero.
The limits, said plainly
You are inside OpenAI’s models: if your work performs better with another lab’s model, you cannot mount it here (that is what OpenCode is for). The quota shared between chat and agent surprises people who use both heavily: the day the agent works hardest is the same day the chat runs out sooner. And as with every agent, quality depends on the task and the supervision: the method of your own cases applies here too.
Testing: descriptive sheet based on official documentation and prices checked on 19/07/2026; no operational procedure included. The hands-on comparison arrives with the coding-agent face-off (ST-07), which requires a ChatGPT account.