About this research · Sources, method & limitations
Check the allowance and reset notice for your own account, then measure one representative task. Include the time spent reviewing and correcting the result. A subscription price or credit balance alone does not tell you the cost of a finished feature. This guide provides a budgeting worksheet, not a verified price per task.
Cost basis: OpenAI's current ChatGPT plan, Codex access, credit and rate-card pages were checked on 25 August 2026. No credits were purchased and no usage dashboard was inspected. Because entitlements and rates can change by plan and account, this worksheet intentionally avoids a permanent “cost per feature” promise.
A useful Codex budget has two ledgers. The vendor ledger records included access, purchased credits and usage. The delivery ledger records planning, review, failed attempts and correction. Ignoring the second ledger can make an apparently cheap task expensive.
A four-step budget check
Use this sequence before quoting repeat work to a client. The example follows a fictional component edit; it is not a measurement of Codex performance or your balance.
Read the allowance
Step 1 of 4Identify the account, usage meter and reset notice relevant to the work. Keep included usage separate from any optional paid balance.
Your proposed job is one contact-form correction. Before starting, record which allowance the account shows; leave unknown values blank.
Stop and review: You cannot tell which allowance or billing route the next task will use.
Define completion
Step 2 of 4Write the permitted change and acceptance checks before measuring usage. Include review work in the task definition.
The form must keep its label visible, explain an invalid email and preserve entered text after a failed save. Deployment is outside this trial.
Stop and review: The task expands into a redesign or unrelated files, making it a different unit of work.
Record actual work
Step 3 of 4Capture observable usage and the minutes spent reviewing, correcting and rerunning checks. Do not count a generated patch as accepted work.
If the failed-save check does not pass, record the repair attempt against this same component task rather than hiding it in a new budget row.
Stop and review: Usage is not yet reconciled, or a check fails without a recoverable correction.
Set the next cap
Step 4 of 4Compare several similar accepted tasks with your normal delivery process. Use that evidence to set a cap with a reserve for rework.
Three contact-form edits can inform another small form task. They do not establish a reliable budget for a database migration.
Stop and review: The estimate depends on unmeasured usage, a different task class or unlimited retries.
Where do I check my Codex limit?
Plan access, included usage and extra credits are different things. Before waiting for a reset, buying credits or changing plans, check the allowance shown for your signed-in account and workspace.
Open the Codex usage dashboard. In the Codex CLI, use /status to inspect the available usage information. Read the meter label, remaining allowance and reset notice together. Do not assume a percentage refers to every limit on the account.
What should I do when a limit is reached?
OpenAI's Codex plan guide, checked on 9 September 2026, describes options that can include waiting for a reset, using credits, redeeming an available reset or upgrading. Availability depends on the account and what the product offers at that time.
Eligible Plus and Pro users can use additional credits after included usage is exhausted. Check the offer and terms in your account before buying. Extra credits do not imply a fixed number of completed tasks.
How should I read a reset notice?
Suppose a meter says Weekly: 30% remaining and resets in 12 hours. That is a hypothetical example, not a reading of your account. The remaining percentage and countdown belong to that weekly meter; another usage window can have a different allowance and reset time.
Use each meter's own label. A short-window reset is not evidence that the weekly allowance has also reset.
How much does one task cost?
There is no dependable fixed price for "one feature." Usage can vary with the model, task complexity, context and tools involved. A short explanation and a multi-file debugging task are not interchangeable units.
Record the account's measured usage alongside review time and rework. The worksheet below helps turn those observations into a budget without promising a universal task count.
Define representative tasks
Price the work you actually repeat:
- explain a failing test without editing;
- change one component and run focused checks;
- trace a bug across several files;
- upgrade a dependency and repair regressions;
- delegate a longer task in an isolated environment.
Run no customer deadline from a marketing allowance. First identify whether your ChatGPT plan includes Codex access, whether extra credits are available to that account and which activity the current rate card measures.
Codex task-cost worksheet
- Plan entitlementvariable
Capture the active plan and the Codex access shown for that account on the test date.
- Measured usageunresolved
Record the dashboard or rate-card unit consumed by one representative task; do not infer it from elapsed time alone.
- Human reviewverified
Time plan review, diff inspection, test interpretation and final acceptance separately from agent execution.
- Reworkverified
Count discarded attempts, reverted files and repeated test runs against the same deliverable.
- Account variabilityvariable
Recheck current credits, limits and rate card before quoting a client or increasing automation.
Separate included use from credits
OpenAI documents Codex access within eligible ChatGPT plans and additional credit mechanisms for some account types. Those are account facts, not universal monthly output. The same plan can produce different useful work depending on task complexity, context, model choice and how often a human redirects it.
For Business, Enterprise or Edu environments, a rate card can express usage through credits. Consumer credit behaviour can use a different help flow. Record the exact page, account and date beside any spreadsheet value so a later pricing change does not silently alter old client estimates.
Add human review and rework
Use this internal formula:
task cost = allocated subscription or credits + reviewer time + rerun time + external compute + rollback cost
Do not publish a precise currency result until every input belongs to your account and market. For a small front-end change, reviewer time may dominate. For a long delegated migration, compute and correction may dominate. A task that finishes quickly but leaves hidden regressions has not reduced delivery cost.
Access update checked on 1 October 2026
OpenAI's plan guide now includes Free and Go in Codex access, but not Codex Cloud. Shared allowances can also cover other eligible apps. This scoped check does not refresh the broader August review or establish a per-task price.
Worked budget example
For a hypothetical component change, suppose task definition takes five minutes and diff and browser review takes twelve. These are illustrative inputs, not observed timings. Add the measured Codex usage shown by your account and any rerun caused by a failed check. Compare that total with the normal manual path. Repeat on at least three similar tasks before using the median for planning.
Do not combine unrelated experiments. A successful documentation edit says little about a dependency migration. Keep a row per task class and include the percentage of attempts that required rollback or substantial rewrite.
Set a budget rule
Choose a monthly cap that includes a rework reserve. Pause the experiment when the account usage is unclear, when reviewer time rises for three similar tasks, or when the task repeatedly crosses into unapproved files. Raise the cap only after the delivery ledger shows a repeatable saving.
The Codex surface selector can reduce handoff overhead; the existing-codebase safety runbook defines the evidence that must exist before a task counts as complete.
No per-task price is verified here because no live account was observed. The worksheet is the original contribution; current plan and credit values must come from the reader's account at the time of purchase.
For a US team or an Indian freelancer billing an overseas client, retain the provider's billing currency in the usage record. Keep your client fee in a separate column and reconcile any conversion or tax using the actual bill. The worksheet does not assume a currency rate or promise identical regional terms.
Revision scope, 9 September 2026: Usage checks and limit-response options were rechecked against OpenAI's Codex plan guide. No credits were purchased or per-task costs measured. The underlying cost worksheet's broader source review remains dated 25 August 2026.
Layout and explanation update, 12 September 2026: Added a step-by-step budget example and clarified the illustrative timings. This revision does not refresh plan entitlements, prices or the fact-check date.
Reader discussion
Comments are public and require a GitHub account. By taking part, you agree to the Terms and Conditions and acknowledge the Privacy Policy.