What counts toward Codex usage?
Usage is driven by the total tokens Codex reads and writes during a session. That includes your prompt, files it opens, terminal output, logs, search results, JSON payloads, tool responses, and the model output it sends back.
This became more literal in April 2026, when OpenAI switched Plus, Pro, and Business plans from message-count limits to token-based metering — so every token of noisy context now maps directly onto how much of your plan you consume.
Which workflows burn usage fastest?
Debugging from logs, codebase exploration, large JSON payloads, shell-heavy tasks, and repeated file reads are usually the worst offenders. These workflows pile on machine-generated context quickly, so a few long sessions can chew through your ChatGPT plan much faster than expected.
How do you get more from your plan?
The best approach is to reduce noise before it reaches the model. Keep the useful context, strip the repetitive stuff, and make verbose inputs smaller before Codex reads them. That is where tools like RTK, Distill, MemStack, and Headroom help most.
If you want the workflow details, read the Codex cost guide and compare the benchmark section against the kinds of sessions your team runs most.
How is Codex usage different from Codex cost?
Usage is about how much of your plan you consume. Cost is about the token spend behind that usage. The two are related, but not identical: the same noisy session that burns through your plan faster also tends to make Codex feel more expensive.
For a deeper look at the limit mechanics specifically, see Codex usage limits and the 5-hour window. For the underlying cost drivers, see why Codex is so expensive.