A Codex message can start many model and tool steps. Count the work needed to finish the task, not just the messages you type. This guide explains consumption; if you cannot start the next task, use the limit recovery steps.
First identify what you are measuring
- Plan allowance: the remaining five-hour or weekly limit displayed for your account. ChatGPT Work and Codex share usage.
- Credits: a billing unit for eligible paid continuation. Published credit rates do not, by themselves, calculate the percentage deducted from included plan limits.
- Context: information a model can work with in a conversation. The context indicator is not a remaining-subscription indicator.
- API billing: a separate usage-based route when signed in with an API key.
See the official pricing explanation, authentication rules and status command. Current plan estimates and rates are collected on the Codex platform page.
Which choices change consumption?
Model and reasoning
A smaller model can suit a narrow edit; an ambiguous debugging task may need stronger reasoning. Higher reasoning effort increases token use according to the subagent and reasoning guidance. It does not establish a fixed percentage cost per message. Inspect the actual selected model and reasoning setting before comparing runs.
Files, tool results and search
Files and tool results add information for the model to process. A short prompt that searches a repository, reads logs and repeatedly tests fixes can involve more work than a long, self-contained question. Check whether searches are targeted and outputs contain relevant lines. No universal “one shell command” or “one web search” deduction is established by the plan documentation.
Cached input
The credit rate card distinguishes regular and cached input. Cached does not mean free, and reopening a file does not prove a cache hit. Codex credit billing has no separate cache-write charge; API-key billing follows its own pricing. Do not convert an API caching discount into a promised subscription saving.
Subagents
Each subagent does its own model and tool work. Parallelism can shorten elapsed time while increasing total token use. Compare the whole completed task, including the main agent's coordination; three agents do not imply either exactly three times the allowance or a guaranteed saving.
Fast mode
Fast trades higher credit consumption for speed on supported models. Check /fast status in the CLI and use /fast off when Standard is adequate. The speed guide separates ChatGPT credit multipliers from API pricing; do not transfer one to the other.
Two examples, not measured savings
For a spelling correction in a known file, ask for that edit and a diff check. A repository-wide investigation or parallel review adds work that the task may not require.
For a bug spanning authentication, storage and UI, independent investigation can be useful. Give each agent a boundary, combine findings, then verify the fix. Faster completion may justify extra consumption; it does not prove lower usage.
Investigate an unexpected drop
Record the account, workspace, model, speed and displayed limits before and after one bounded task. Check concurrent Codex and ChatGPT Work activity, then inspect repeated reads, retries and subagent work. Change one factor at a time. This is a troubleshooting method, not a controlled benchmark or a formula for future allowance.
An exhausted limit still needs the account's reset time or an eligible continuation option. Reducing future work does not refund past usage.