ChatGPT Project File Limit and How to Organise a Big Folder

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for ChatGPT Project File Limit and How to Organise a Big Folder.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for ChatGPT Project File Limit and How to Organise a Big Folder.

A ChatGPT Project holds 5 files on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise and Edu, uploaded at most 10 at a time. For a big business folder, merge related documents into fewer, well-named files, split the work across several projects by job, and ask every answer which file it used.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

The file count is the limit that bites. Each document can be up to 512 MB, and text files are capped at 2 million tokens (OpenAI's rule of thumb is a token for about three-quarters of a word, so roughly 1.5 million words), which no ordinary contract or handbook comes near. Projects themselves are unlimited. So the job is not squeezing a folder into one project; it is deciding which documents belong together, and in which project, so that answers come from the right file.

Consider an illustrative courier firm running nine vans, with a shared drive folder of 212 files built up over years: customer contracts, rate cards, service-level schedules, the driver handbook, health and safety procedures, vehicle records, claims procedures and past tender submissions. The office manager wants staff to ask ChatGPT questions such as "what is the waiting-time charge for this account?" and get an answer with a source, on a ChatGPT Business plan.

If you have not set up a project before, start with customising ChatGPT with Projects, instructions and files, which covers creating one and writing instructions. Once a project exists, the limits and the way you organise its files decide whether it gives good answers.

What are the ChatGPT Project file limits on each plan?

ChatGPT Projects cap the number of files, not their pages: 5 on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise and Edu, with each document up to 512 MB.

Project limits come from two OpenAI help pages (Projects, and file uploads) and are worth having in one place before you plan anything.

PlanPriceFiles per projectCollaborators on a shared project
Free$055
Go$8 a month2510
Plus$20 a month2510
ProFrom $100 a month40100
Business$25 per seat monthly, $20 billed annually40Up to 100 on a workspace project
Enterprise, EduNot on the pricing page40Up to 100 on a workspace project

The limits that apply to every file, whatever the plan:

  • Documents and presentations: up to 512 MB per file; text and document files capped at 2 million tokens each.
  • Spreadsheets and CSV files: roughly 50 MB, depending on row size. The token cap does not apply to spreadsheets.
  • Images: 20 MB each.
  • Upload pace: up to 80 files every 3 hours; Free accounts are limited to 3 file uploads a day; only 10 files can go into a project at once. OpenAI may lower limits at peak times.
  • Storage: 25 GB per user and 100 GB per organisation for shared upload storage.
  • Linked sources: Slack channels and Google Drive files added as project sources have a separate limit from uploads (OpenAI does not publish the number).

The upload pace matters more than it looks. On Free, the courier's 44 organised files would take 15 days to upload at three a day, even before the five-file project limit forced further merging. On a paid plan, a 212-file dump would hit the 80-files-every-3-hours cap long before it hit the point that no single project could hold it. Organising first means uploading far fewer files.

Why do fewer, larger files beat many small ones?

Fewer, larger files beat many small ones because a project counts files, not pages, so merging is how a courier firm's 212 documents fit inside a plan. A project with 40 tiny files and a project with 40 merged files have the same count, but the second holds far more of your business.

There is a limit to how far to take it. OpenAI notes that large, complex or image-heavy files may not be fully analysed, and that very large workbooks may not fit within the context window, which can lead to partial results. A single 900-page "everything" PDF is technically allowed and practically a bad idea: answers become harder to trace, and the file is painful to replace when one page changes. The rule that works is one topic per file: merge documents that answer the same kind of question, and keep topics apart.

Scanned paperwork needs its own warning. OpenAI says ChatGPT may not reliably extract exact values from scanned files or image-based tables. The courier's older customer contracts were scans; the office manager re-saved them as searchable PDFs, using the text recognition (often called OCR) in the office's scanning software, before merging, then spot-checked three clauses in each against the paper copy.

Step 1: How do you sort a big folder before uploading it?

Sort a big folder by the questions your team will ask ChatGPT, not by department or file type, because a project should answer one family of questions well. The courier's audit, filled in:

Folder contentsFilesTypical questionDecision
Customer contracts46"What's the waiting-time charge for account RT-114?"Merge into 6 files by customer segment
Rate cards18"What do we charge this account for a 30 kg next-day parcel?"Merge into 1 workbook, one tab per account plus Standard
Service-level schedules12"What's our on-time commitment to this account?"Merge into 1 file
Driver handbook and policies25"What must a driver do if a parcel is damaged?"Merge into 4 files by topic
Health and safety procedures30"What's the manual-handling limit per lift?"Merge into 5 files by topic
Vehicle records and maintenance logs34"When is van 7 due a service?"1 log workbook plus 5 reference files
Claims and incidents22"How do we handle a lost-parcel claim?"Keep 2 procedure files; leave individual incident reports out
Past tender submissions15"How did we describe our tracking in past tenders?"Keep all 15 separate, plus 2 reference files
Miscellaneous10None in particularDelete 6 obsolete files; 4 join other topics

Two decisions in that table deserve a reason. The individual incident reports stay out because they name drivers and customers, and a shared project shows its files to every member; the procedures answer the "how do we handle this?" questions without exposing anyone. The tenders stay as separate files because the useful answer is "in the March tender for a pharmacy group, we said X", and a merged file would blur which tender said what.

Step 2: How do you merge documents without losing the trail?

To merge documents without losing the trail, make sure any answer drawn from the merged file can be traced back to the original document. Three habits keep the trail:

  1. A contents page at the front. List each original document, its date and where it starts. For the contracts, that meant account code, customer segment, contract date and page number.
  2. A heading at the start of every original. "ACCOUNT RT-114: CONTRACT AND SCHEDULES, signed [date]". Clause numbers repeat across contracts (every one has a clause 4.3), so the heading is what tells ChatGPT, and you, whose clause 4.3 it is.
  3. One topic per file, and nothing superseded. If a rate card was replaced, the old one is deleted, not merged in. Two versions of the same price in one project is the commonest source of confidently wrong answers.

The rate cards went the other way: 18 small workbooks became one workbook with a tab per account and a Standard tab, each tab with the same column layout (service, weight band, price, surcharge). Identical layouts matter; when every tab has its columns in a different order, questions that compare accounts go wrong.

The merge that went wrong first time

The first merged contracts file had no headings, only the original documents joined end to end. Asked for the waiting-time charge for account RT-114, ChatGPT answered "$18.00 per 15 minutes after the first 20 minutes" and cited the merged file, clause 4.3. The citation was real, but it was the clause from the contract on the pages before, for a different account; RT-114's own clause 4.3 says $15.00 after 30 minutes. Nothing in the answer looked wrong. The office manager only caught it because RT-114 was one of the accounts in the test set (Step 5). Adding the account headings and the contents page fixed it, and the instruction to quote the account heading with every contract answer made the source visible.

If you want help fitting this into how your office actually runs, I offer one-to-one AI consultations for small businesses.

Step 3: How should you name files in a ChatGPT Project?

Name each file so it tells a reader what it is, which version it is and whether it is current, because answers and the project's sources list both show file names. The courier's convention was TOPIC_Subject_date, with the date as year-month so files sort in order:

BeforeAfter
Scan_0047.pdf, Scan_0052.pdf and 4 moreCONTRACTS_Pharmacy-segment_2026-10.pdf
rates FINAL v2 (3).xlsx and 17 moreRATES_All-accounts_2026-11.xlsx
Handbook new.docx, Damage procedure.docxDRIVERS_Damaged-and-lost-parcels_2026-09.docx
Tender - Copy.pdfTENDER_Pharmacy-group_2026-03.pdf
(none)INDEX_What-is-in-this-project_2026-11.md

The INDEX file is the cheapest improvement of all: a one-page list of every file in the project, what it covers, its date and who owns it. Staff read it to know what the project can answer; ChatGPT can use it to know where to look. Update it whenever a file changes.

If the spreadsheet side is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

Step 4: Should you split the work into several projects?

Yes, split the work into one project per job, because shared projects show every file to every member and each job has its own people and questions. With 44 files left, the courier could have squeezed most of them into one 40-file Business project. It made four instead.

  • Customer contracts and pricing (9 files: 6 contract files, the rate workbook, the service-level schedules, the index). Shared with the owner and two office staff. Instructions: "Answer only from these files. Name the file and quote the account heading and clause for every answer. If the answer is not in the files, say so."
  • Operations and drivers (12 files: 4 driver files, 5 health and safety files, 2 claims procedures, the index). Shared with the operations lead and the two supervisors.
  • Fleet (6 files: the maintenance log workbook, 4 reference files and the index). Used by the owner and the fleet supervisor.
  • Tenders (17 files: the 15 past submissions, a company-facts file and a boilerplate-answers file). Used by the owner when a tender comes in.

Every project sits under 25 files, so this layout would also work on Plus. That is a useful test of any plan: if your layout only works on the biggest plan, you have probably merged too little or mixed too many jobs in one project. The principle is the same one that keeps client work separate in ChatGPT Projects: one project per job, so nothing leaks into an answer that should not be there.

The Fleet project behaves differently from the others because its main file is a spreadsheet. "When is van 7 due a service?" is a calculation, not a look-up: find van 7's last service date in the log and add the service interval from the reference file. Ask for the method as well as the answer ("show the last service date you found and the interval you added"), so a misread date shows up at once. Keep the log trimmed to the columns that matter; OpenAI warns that very large workbooks may not fit in the context window and can produce partial results.

Making it work on Free's five files

Free allows 5 files per project and 3 uploads a day, which changes the plan but not the method. A sole trader running deliveries could fit the essentials into one Free project with five merged files: CONTRACTS (all customers, with account headings), RATES (one workbook), PROCEDURES (driver, safety and claims procedures together, each section headed), FLEET (the log) and INDEX. Uploading takes two days at three files a day. Answers will be slower to trace than in the courier's four-project layout, because each file covers more ground, so the instruction to quote the heading and clause matters even more.

Three settings to know when you share. Shared projects are switched to project-only memory automatically and cannot be switched back, so chats inside them do not draw on anyone's personal memories. Project members can see the project's chats, files and member list. And on Business, workspace owners control who can create projects through the Create and manage projects permission under Workspace settings, then Permissions & roles.

Step 5: How do you check which file each answer used?

Check which file each answer used by telling the project to name the file, heading and quote every time, and by running a known-answer test whenever the files change. An answer without a source is a guess you cannot audit. Two checks keep the project honest:

The habit is built into the project instructions and repeated in the question when it matters:

For account RT-114, what do we charge when a driver waits more than
30 minutes at collection? Quote the exact sentence, and give the file
name, the account heading and the clause number you took it from.
If the files do not say, answer "not in the project files".

An illustrative answer after the fix:

$15.00 per started 15 minutes after the first 30 minutes of waiting. Source: CONTRACTS_Retail-segment_2026-10.pdf, "ACCOUNT RT-114: CONTRACT AND SCHEDULES", clause 4.3: "Waiting time beyond 30 minutes is charged at $15.00 per started 15-minute period."

Then check it: open the project's sources list, preview the file, find clause 4.3 under the RT-114 heading, and confirm the wording. Ten seconds when the answer names its source; impossible when it does not.

A known-answer test you can repeat

Write ten questions whose answers you already know, each with the file that should supply it, spread across the project's topics. Run them after any reorganisation and score two things: was the answer right, and did it cite the right file? The courier's contracts-and-pricing project, scored (illustrative):

Version of the projectCorrect answersCorrect file citedWhat went wrong
First try: 25 files uploaded as they were6 of 105 of 10Old rate card answered two questions; one scan unreadable; one answer had no source
Merged, no headings8 of 107 of 10The RT-114 clause 4.3 mix-up; one weight band read from the wrong tab
Merged, headings, index, old versions removed10 of 1010 of 10None

The weight-band miss came from one tab in the rate workbook with its columns in a different order from the rest. It was fixed in the workbook, not in ChatGPT, which is typical: most wrong answers in a well-instructed project are document problems, not model problems.

How do you keep a ChatGPT Project current?

Keep a ChatGPT Project current by replacing changed files, keeping masters elsewhere and re-running your test questions, a routine that takes a few minutes a month. A project is a working copy of documents that keep changing. The steps:

  • Replace, never add. When a document changes, delete the old file from the project's sources list (where uploaded files can be previewed, downloaded or deleted) and upload the new one with the new date in its name. Then update the INDEX file.
  • Keep masters elsewhere. Deleting a project permanently removes its chats, instructions and the files stored only in it, and cannot be undone; files saved separately in Library or Space survive. Keep the master copies in your own shared drive, so any project can be rebuilt in half an hour.
  • Save good answers back in. When a chat produces a useful summary (a one-page comparison of three accounts' surcharges, say), open the response's message menu and choose Add to project sources, then rename it in your index.
  • Treat Drive links with care. Linking a Google Drive file avoids using an upload slot, but OpenAI notes the Google Drive app does not support sync when added within a project. Re-run a known-answer question after you change a linked file, and upload a fresh copy if the answer is stale.
  • Re-run the test quarterly. Ten questions, ten minutes. A score that slips from 10 to 8 is usually a superseded file someone forgot to remove.

If projects ever vanish from the sidebar after a workspace change, the tutorial on where missing ChatGPT Projects go explains the usual causes. And if your folder is mostly client files rather than internal documents, Claude Projects compared with ChatGPT Projects for client files is worth reading before you commit to one tool.

ChatGPT Project file limits: questions owners ask

Does a linked Google Drive file count towards the project's file limit?

Not towards the uploaded-file limit. OpenAI says Slack channel and Google Drive links added as project sources have a separate limit, though it does not publish the number. It also notes the Google Drive app does not support sync when added within a project, so check that answers reflect the current version of a linked file before relying on them.

Can I get more than 40 files into one project?

No, not by uploading more: 40 is the ceiling on Pro, Business, Enterprise and Edu. The practical route is fewer, larger files: merge documents that answer the same kind of question into one file with clear headings, and move unrelated material into its own project. Projects themselves are unlimited, so splitting costs nothing except a moment's thought about which project to open.

What happens to the files if I delete a project?

They are removed for good if they were stored only in that project. OpenAI says deleting a project permanently removes its chats, instructions and the files stored only in it, and the action cannot be undone. Files you also saved separately in Library or Space stay there. Keep your master copies in your own shared drive, so the project is always a working copy you can rebuild.

Why does ChatGPT sometimes answer from the wrong file?

Usually because two files say similar things, such as an old and a new rate card, or two contracts with the same clause numbers. Remove superseded versions, put a clear heading at the start of each merged document, and tell the project instructions to name the file and section for every answer, so a wrong source shows up immediately.

Further reads

Sources: OpenAI help pages on Projects in ChatGPT, file uploads, data analysis and tokens; the ChatGPT pricing page; OpenAI's ChatGPT for Excel and Google Sheets page, all checked 17 November 2026.

Have a specific process in mind?

Describe the job in your own words on a 1:1 call and I'll show you what AI you can use for it with the tools you already pay for.

Book a 1:1 call with me