You can make your Codex and ChatGPT Work allowance last longer by reusing prepared files instead of asking the agent to find and clean them again. Save the original reports, a usable table and a quick handoff note so the next question can start from there.
Imagine this: yesterday, the AI pulled six sales reports into one tidy table. Today, you only want to include returns in the calculation. It starts hunting for the six reports again. Where did yesterday’s table go?
One job, two very different costs
Codex handles code and technical changes. ChatGPT Work handles multi-step work involving files, websites and tools. OpenAI says Work and Codex draw from the same usage allowance included with your plan. How much usage and credit each plan includes can differ. How much a task uses depends on the model, the work involved and the size of the input and output. Where available, reasoning level and speed settings matter too.
OpenAI notes that GPT-6 Astra may use the allowance faster than GPT-5.6 Sol. Some plans have both 5-hour and weekly limits; Settings → Usage shows the remaining allowance and reset times. As of October 2026, Pro $100, $200 and $500 plans have no 5-hour Work/Codex limit, although their included usage allowance still applies. On Pro $500, Ultrafast draws from the included allowance at 8 times the Standard rate. After that allowance is exhausted, it uses credits at six times the Standard credit rate. Check the speed setting before blaming repeated file setup.
Before blaming repeated downloads for a shrinking allowance, check whether a more intensive or faster mode is selected. Regular Chat has a separate usage structure.
Model choice is another lever. If GPT-6 Luna is available on your plan, try it for renaming columns or tidying a short checkpoint; reserve a stronger model for a hard bug or unfamiliar system. Check the result before calling the cheaper option a win: a failed first pass followed by three retries is hardly a bargain.
Imagine importing six months of sales records. Connecting to the source, handling inconsistent dates and checking missing rows is preparation. Asking whether sales improved after a price change is analysis. If you prepare the same records for every question, you've effectively paid for the kitchen to be rebuilt every time you order lunch.
Save what your next question needs
After the first pass, keep three things: the original reports (with their source and retrieval date), a cleaned table (with column definitions and filters), and a checkpoint (what was checked and what remains uncertain). If you may need to recalculate, keep the formula you used too.
Keeping six original reports does not mean sending all six back to the agent for every question. Start with the cleaned table and checkpoint; retrieve only the source records needed for the new calculation. Bigger inputs can consume more allowance, so reuse saves work when it actually reduces repeated reading.
There is a catch: if the first table excluded returns, it cannot magically answer a returns-inclusive question. Go back to the saved originals, include those records and recalculate. Reusing preparation is not the same as skipping a necessary calculation.
Next time, begin with the prepared table and checkpoint. If returns were filtered out, open the relevant records in the saved originals. Then ask: “Add the return records from the saved originals to the prepared sales table, recalculate net sales, show which totals changed, and check three refunds against their source entries.” That is a new calculation built on saved preparation, not another hunt for the same six files. Ask Chat to compare scenarios or explain the results. If the source has changed, the data is too large or another tool is needed, return to Work, Codex or the appropriate application for that part. Don't pretend Chat can access a system it hasn't been given.
The same principle applies to coding: keep the test results and current error, not just “the program still fails.” A useful checkpoint saves the next person—or the next AI—from detective work.
Keep the savings, check the result
If returns change the total, spot-check a few orders against the original reports. Were refunds subtracted rather than counted as new sales? Avoiding a repeated download saves time; trusting a wrong total can cost much more.
Before your next session, check that you can actually open the source file, the cleaned table and a short progress note. Reuse those when only the question changes; use Work, Codex or another tool when the data or the outside world really needs updating. Your exact allowance will still depend on your plan and the task.
The most useful last request might be: “Save this so I won’t have to ask you to do it again.” Asking an assistant to make itself less necessary is a surprisingly good use of an assistant.
References (3)
OpenAI Help Center, Using Codex with your ChatGPT plan.
OpenAI Help Center, ChatGPT Work and Codex.
OpenAI Help Center, Managing usage with GPT-6 Astra in Work and Codex.
