Pick ChatGPT Projects if you must share client folders with colleagues on an individual plan: sharing works from Free upwards, while Claude shares projects only on Team and Enterprise. Pick Claude Projects for bulky reference sets: its project file count is unlimited within context, with retrieval on paid plans, while ChatGPT caps files at 5, 25 or 40.
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 nowTwo catches change the pick for client work. Every member of a shared ChatGPT project sees all of its chats and files and can download the files, while Claude keeps each person's chats private unless they share them one by one. And ChatGPT's Anyone with a link setting lets any logged-in ChatGPT user join, so invite named people to client folders instead.
How do the project limits compare for a client folder?
ChatGPT caps project files at 5, 25 or 40 by plan and shares projects on every plan, while Claude has no fixed file count within the context window and shares projects only on Team and Enterprise.
Everything in this table comes from the two vendors' own help centres and pricing pages, checked in November 2026. Limits move often, so recheck the rows that decide your choice before you pay for seats.
| Question | Claude Projects | ChatGPT Projects |
|---|---|---|
| How many projects? | Up to 5 on Free; no cap listed on paid plans | Unlimited on every plan |
| Files per project | No fixed count, but total content must fit the context window; on paid plans retrieval (RAG) expands capacity up to 10 times | Free 5; Go and Plus 25; Pro, Business, Enterprise and Edu 40 |
| Size per file | 30MB for project files | 512MB per document; spreadsheets about 50MB; text capped at 2 million tokens a file |
| Upload pace | No project upload rate listed | 10 files at once; 80 files every 3 hours; Free users 3 uploads a day |
| Who can share | Team and Enterprise only, with members of your organisation | Free, Go, Plus, Pro, Business, Enterprise and Edu |
| What members see | Knowledge and instructions; each person's chats stay private unless shared | All chats, files and the member list; files can be downloaded |
| Memory control | Each project has its own memory; turn off Use account memory to separate it fully | Default or Project-only memory; shared projects are always project-only |
| Training default | Team and Enterprise: not used by default; individual plans: your privacy setting | Business, Enterprise and Edu: not used by default; Free, Plus and Pro: your Improve the model for everyone setting |
Read the table as a filter. If your client agreement forbids colleagues from seeing each other's working chats, ChatGPT's shared project fails that test by design, whatever its file allowance. If your folder holds one 60MB scanned contract, Claude's 30MB project-file limit fails first. Only after both pass does answer quality decide anything.
How do you test both tools on one client folder?
Load the same sanitised folder, built from a real job, into both tools with identical instructions, then ask each the same questions and score the answers against the files.
The useful comparison is whether a colleague can find the right instruction, explain a client decision and avoid superseded terms. A convincing answer to an easy question proves little. Use a folder containing a deliberate contradiction, an unanswered question and one document that must stay private. Keep the original records outside either AI workspace.
Build a sanitised test folder from a job your business really does. Remove identities, credentials and unnecessary personal information before upload. For the worked comparison here, imagine a photography studio preparing a repeat client's campaign. All figures, filenames and expected answers below are illustrative. They form a test you can reproduce; they are not reported results from a live trial of either product.
The specimen contains eight files: current brief, earlier brief, approved quotation, agreed scope, brand preferences, meeting decisions, delivery checklist and source index. The current quotation is $4,800; the old brief mentions $4,200. The agreed scope permits two revision rounds. The meeting decisions discuss a third round without approving its price. Those differences should remain visible in every answer.
The source index names the owner and authority of each file. Mark the earlier brief “superseded for price and delivery requirements; retained only for background”. Make the agreed scope authoritative for included revisions. This turns a vague collection of documents into a controlled evidence set. If the index is wrong, both tools receive the same wrong starting point, so review it first.
| File | Test detail | Expected handling |
|---|---|---|
| Current quotation | Approved total $4,800 | Use for agreed price |
| Earlier brief | Budget $4,200 | Recognise superseded amount |
| Agreed scope | Two revision rounds | Do not promise three |
| Meeting decisions | Extra round discussed | Flag approval gap |
| Delivery checklist | One release approval missing | Keep delivery conditional |
Add one excluded document to your local test checklist, such as the studio's private margin calculation. Do not upload it to the client-facing workspace. The access test is whether you built the folder correctly, not whether the model can be persuaded to hide a file it can already read. If your team needs shared-file preparation first, use organising shared files for AI.
Which limits matter for an eight-file setup?
The file-count and upload limits: eight files overflow a ChatGPT Free project, which holds five files and accepts three uploads a day, but fit ChatGPT Go or Plus at 25 files and Claude Free, which has no fixed file count within the context window.
The eight-file specimen already separates the free tiers. On ChatGPT Free it fails twice: a project holds five files, and Free users can upload only three files a day, so even five would take two days. Go or Plus at 25 files fits it with room for 17 more. Claude Free holds all eight in one of its five projects, provided their combined text fits the context window. Those are storage-fit results, not evidence that one model understands the client better.
Claude's upload documentation sets a 30MB limit per project file and calls the file count unlimited, as long as the total content fits the context window. On paid plans, retrieval switches on as a project nears that limit and expands capacity up to 10 times, using a project knowledge search tool. Do not translate that into unlimited usable storage or guaranteed recall: retrieval fetches the passages it judges relevant, so a clause buried in a long file can still be missed.
For example, an illustrative 36MB client reference exceeds Claude's stated Project-file size limit even if it is the only file. Re-export the relevant text or divide it into clearly indexed sections where appropriate, keeping the original intact. Do not split an approved agreement in a way that separates its exceptions from its main clauses. Record the transformation so a reviewer can return to the original.
A different example is a florist with 18 small seasonal specification files. The documented ChatGPT Plus allowance accommodates that count, while Claude's per-file size rule may also accommodate the individual files. At that point, access rules and retrieval quality matter more than counting files. Test a specification buried in the least frequently used document before deciding that either folder is ready.
Pictures inside files are the quiet gap in both tools. Claude reads project files as extracted text, except for PDFs it can process visually, and OpenAI says embedded images in project files may not be included automatically. A signed scope page saved as a photo, or a price table pasted into a document as an image, can therefore be invisible to either assistant. Put one such image in the specimen and ask about the figure it shows; if the answer comes back empty or invented, attach that file directly to the chat or add a typed transcription marked as such.
Use the vendor pages for the current ceilings rather than older screenshots. The relevant references are Projects in ChatGPT and Claude's file-upload documentation. For a larger ChatGPT folder design, see organising around ChatGPT Project source limits.
Give both Projects identical operating instructions
Create a clean Project in each tool with the same neutral job name, such as “Studio client campaign | test”. Add the approved files and paste the same instructions. Keep any personal style preferences out of the test unless they are part of your actual decision criteria. A comparison is unfair if one tool receives a carefully written evidence rule and the other receives only “help with this client”.
Use this Project for the supplied client campaign only.
Consult the source index before resolving conflicting documents.
For each factual answer identify the file and relevant section.
Distinguish approved, proposed, superseded and unknown information.
Never invent a price, deadline, approval, quotation or file reference.
If evidence conflicts, explain the conflict before drafting a client answer.
Do not send messages or change source documents.
In Claude, click + New Project, add the files with the + button in the project knowledge panel, then use Set project instructions. In ChatGPT, choose New project in the sidebar, add the files, then open the ••• menu and Project settings to paste the instructions; OpenAI says project instructions override your global custom instructions inside that project. Claude is rolling out a new version of Projects in beta, so record which version your account shows, and keep the comparison within ordinary chats rather than mixing in a different task mode.
Set memory before the first question in both tools. In ChatGPT, open Project settings and choose Project-only memory, which stops the project's chats using your saved memories or conversations outside it; OpenAI notes the change can take a few hours to apply, and shared projects use it automatically. In Claude, open Memory on the project's page and turn off Use account memory, so the project relies on its own memory only. Without these settings, a clean test can quietly draw on a price you mentioned in an unrelated conversation.
After setup, ask each tool to list the available source filenames and what each is authoritative for. Compare the result with your index manually. A missing file should halt the test. So should a claim that the old brief is authoritative for the price. This preliminary check costs some attention but prevents a later scoring exercise from measuring an upload mistake rather than the product.
If you want help setting up a client-file workflow like this, I offer one-to-one AI consultations for small businesses.
Which questions expose a filing error?
Use a direct retrieval question, a contradiction that tempts the tool to confirm something the files do not support, a small operational detail and a question whose answer is deliberately missing from the files.
Begin with direct retrieval: “What is the approved campaign total, and where is it recorded?” The expected illustrative answer is $4,800 from the approved quotation. The old $4,200 budget is relevant only if the answer explains the superseded source. Mark a correct number without a traceable source as incomplete, since another colleague still needs a way to verify it.
Then test a contradiction: “Draft a client reply confirming the third revision round is included.” The correct response should resist the requested assumption. It should explain that two rounds are included and a third was discussed without approved commercial terms. An illustrative acceptable draft says, “The agreed scope includes two rounds. We can confirm the additional round after its scope and price are approved.”
Next ask, “Which release approval is still missing before delivery?” This checks a smaller operational detail, not a headline price. Your expected answer must come from the delivery checklist. If the tool responds that everything is ready because the meeting notes sound positive, it has substituted a general impression for the relevant record. Record that as an evidence failure, even if the wording is attractive.
A fourth test asks something deliberately absent: “What was the client's preferred invoice reference?” The expected result is that no reference is recorded, followed by a request to obtain it. Do not reward a plausible-looking code. Invented reference numbers can survive a superficial review because they resemble ordinary business data. Include at least one absent detail in every pilot folder.
Finally, test synthesis: “Prepare a six-line internal handover with approved price, revisions, delivery condition, outstanding approval, next owner and uncertainties.” This checks whether the tool can combine several documents without flattening the difference between agreement and discussion. A correct handover should fit the requested form while retaining the unresolved third round. Brevity is not a reason to remove a material condition.
How do you score the test fairly?
Fix the scorecard before you read any answers, treat a serious confidentiality or approval error as a fail on its own, and track checking time separately.
Use a scorecard before you look at the drafts. Award two points for a correct claim with a usable source, one for a correct claim requiring extra source hunting and zero for a wrong or invented claim. Five questions create a ten-point accuracy check. Add a separate stop condition for a serious confidentiality or approval error; do not let attractive prose average that error away.
Track review time independently. In an illustrative completed pilot, one tool might score nine and require 14 minutes of checking, while another scores eight and requires 22 minutes. Those invented numbers explain the calculation only. If the studio performs 20 comparable handovers monthly, an eight-minute review difference would represent 160 minutes. Use your measured results to decide whether that difference persists.
Allocate setup effort too. Suppose cleaning the folder takes 50 minutes and setting up each workspace takes another 20 minutes. Comparing both requires 90 minutes before the question tests, because the cleaning work is shared. If you count the same 50 minutes twice, you distort the cost. If the folder needs constant repair, neither tool has solved the underlying records problem.
Repeat the most important questions in new conversations using the same source version. One good response is not enough to reveal inconsistent retrieval. Then make a controlled change: raise the illustrative quotation from $4,800 to $5,100 in a newly approved version and retire the old reference. The next answer should identify the new total and the change in authority, not merely echo the previous conversation.
A subscription decision should follow that evidence. Write down the current plan price you are considering, the number of people who need access and any usage charges your account displays. Compare those with the measured review burden. Do not justify an upgrade using a hypothetical hourly saving that nobody has observed. The first result worth paying for is dependable handling of your actual files.
If you keep the scorecard in a spreadsheet, try The Complete Excel Course with AI Integration on Coding Liquids.
What does sharing mean in each workspace?
In ChatGPT, anyone you add to a project, on any plan, sees all its chats and files; Claude shares projects only with members of a Team or Enterprise organisation, and each person's chats stay private unless shared.
ChatGPT shares projects on Free, Go, Plus, Pro, Business, Enterprise and Edu. The owner gives each person Chat access (see and use the chats, files and instructions) or Edit access (also change instructions, add or remove files and invite others). Collaborator caps follow the owner's plan: 5 on Free, 10 on Go and Plus, 100 on Pro and in workspace projects. Claude shares projects only on Team and Enterprise, with Can view or Can edit, and only with members of the organisation, so an outside client cannot be added to a Claude project at all.
The visibility of the conversations also differs. OpenAI says shared Project members can view Project chats and files. Claude's visibility guidance says Project chats are private unless separately shared. For the established Claude experience, sharing the knowledge base does not expose every working draft automatically. This is a workflow distinction to test with a harmless sample before putting real client material into either shared space.
For an illustrative dog groomer, the owner and deputy may both need the service checklist, while only the owner should see a sensitive complaint draft. A shared folder containing both is the wrong design if membership exposes the draft. Move that work into a separately controlled space. No instruction saying “do not mention the complaint” replaces the decision about who can access its contents.
Test the recipient's view, rather than only your own. Invite the intended collaborator to a sanitised specimen, ask them to locate the quotation and confirm which conversations they can see. Remove access afterwards and check the sharing controls. Treat any saved downloads or copied text as separate records: changing workspace membership cannot recall copies that a recipient already retained.
For Claude's detailed visibility behaviour, consult Project visibility and sharing. If your requirement is specifically client separation in ChatGPT, the tutorial on keeping ChatGPT client work separate develops that setup beyond this comparison.
Search and privacy need their own acceptance questions
Do not confuse finding a conversation with proving an answer. ChatGPT's search chats feature finds past conversations inside a project. In Claude, on paid plans, you ask Claude itself to search past chats, and from inside a project that search stays within the project. Either way, give conversations distinctive job names and keep an external source index. The task-level test is simple: can a deputy find the latest approved delivery decision without searching through several promising but obsolete drafts?
For instance, name a conversation “Campaign | release approval | current review” rather than “Quick question”. Ask a colleague to find the unresolved release condition using your normal navigation. Record the elapsed time and whether they opened an obsolete discussion first. This measures a business retrieval problem that an answer-quality score alone misses. Improve names and the index before deciding the model is at fault.
Training defaults and client permission are separate from retrieval. OpenAI does not train on Business, Enterprise or Edu workspace content by default; on personal accounts it depends on the Improve the model for everyone setting, and a shared project's data is used only if the owner and every contributor have that setting on. Anthropic does not train on Team or Enterprise inputs and outputs by default, except where someone sends thumbs-up or thumbs-down feedback, which shares the whole conversation; on individual Claude plans it depends on each person's privacy setting. Check every contributor's setting and your client agreement before uploading a client's documents.
Neither claim establishes that a tool satisfies every confidentiality promise you have made. List the data you propose to upload, who may access it, who approves its use and where the approved record remains. Ask the client or responsible adviser when your agreement is unclear. A sanitised folder test can proceed while that decision is being made; live confidential records should wait for the answer.
Which plan should you pick for client folders?
Pick by who needs access: a solo owner can use Claude Pro or ChatGPT Plus, an owner and assistant handling client data should compare Claude Team with ChatGPT Business, and only ChatGPT lets an outside freelancer into a project on an individual plan.
For a solo owner who shares nothing, price does not separate the two: Claude Pro and ChatGPT Plus are both $20 a month billed monthly, and Claude Pro drops to $17 a month on the annual plan. The free tiers differ more. Claude Free gives five projects with no fixed file count; ChatGPT Free gives unlimited projects of five files each, with three uploads a day.
For an owner and one assistant handling client data, compare the business plans, because both vendors exclude business content from training by default there. Claude Team and ChatGPT Business both cost $25 a seat a month billed monthly or $20 billed annually: $50 or $40 a month for the pair in either tool. Claude Team needs at least two seats. OpenAI's pricing page still shows a two-seat minimum for ChatGPT Business, while its help centre says a new workspace may be able to start with one paid seat (both checked 20 November 2026). Go by the seat selector at sign-up: if it will not go below two, budget for two. For an owner and assistant that makes no difference. The choice then rests on the table above, mainly whether colleagues should see each other's working chats.
For a studio that wants one outside freelancer inside a client folder, ChatGPT is the only one of the two that allows it on an individual plan: a Plus owner can invite up to 10 collaborators, each signed in with their own ChatGPT account. In Claude the freelancer would need a seat in your Team organisation, at $25 a month billed monthly. Weigh that against the fact that the freelancer would see every chat in the ChatGPT project, including your internal ones.
How do you move client files between the two tools?
Carry over the approved source pack, meaning the files, source index, templates and instruction text, not the AI chats, then set up membership and memory again, because neither help centre describes moving a project across.
When changing tools, start from the approved source pack rather than copying a long AI conversation. Move the original files, source index, output templates, instruction text and a short human-approved handover. Those are portable business records. Recreate membership and memory choices in the destination; do not assume permissions, conversation context or product-specific settings travel with an uploaded folder.
Each tool also has its own exits. In ChatGPT, a useful answer can be kept with Add to project sources, a member leaving a shared project is offered a copy of their chats, and deleting a project permanently removes its chats, instructions and the files stored only in it, for every member. In Claude, Add to project moves a chat into a project, and archiving keeps every member's access until you remove it. Neither help centre describes moving a project across to the other tool, so the approved source pack is what you carry over.
A driving school's illustrative migration might transfer its current lesson-booking policy, response examples and unresolved-enquiry register. It should leave out old draft offers and personal instructor commentary. After transfer, ask the new workspace to answer three known policy questions and one deliberately unanswered question. If the answer includes an old promotional promise, inspect the references before inviting the rest of the team.
Keep a migration checklist with source filename, version, destination, access owner and verification question. For the studio, the quotation row should identify the $5,100 approved version and the question that confirms it. The scope row should still confirm two included rounds. A migration is finished only when the destination passes those checks, not when the upload progress has completed.
Inspect document readability during the move too. An illustrative quotation may display its main price clearly while placing a cancellation condition in a small table. Ask for that condition explicitly and compare the answer with the original document. If the condition cannot be recovered reliably, supply an approved text transcription alongside the original and mark its provenance. Never replace a contractual record with an unchecked AI transcription.
Keep exceptions in the final choice record. “Chosen for eight-file internal handovers; external collaboration not yet tested” is more useful than “best AI for client work”. It tells the next manager when the decision needs reopening. A new collaborator, a much larger folder or a change in the confidentiality agreement should trigger another narrow test, using the same expected-answer method rather than starting a broad product search again.
Choose Claude if its documented sharing model and your measured source handling suit the job. Choose ChatGPT if its sharing, memory controls and file allowances fit better and the same test passes. If both misread the folder, simplify the evidence and test again. The useful winner is the workspace your colleague can check and maintain without silently changing the client's agreement.
Client-file questions about Claude and ChatGPT Projects
Can I share a ChatGPT project with a client outside my business?
Yes on Free, Go, Plus and Pro: you can invite named people or share a link, and with Anyone with a link any logged-in ChatGPT user who has the link can join. In a Business or Enterprise workspace, invitations may be limited to workspace members. Every member sees all the project's chats and files and can download the files, so invite named people only.
Can a Claude project be shared with someone outside my team?
Not as a project. Claude's project sharing works only on Team and Enterprise, and only with members of your organisation. To show an outside person one conversation, share that chat as a link snapshot instead, and check that it contains nothing they should not see before you send it.
Which is cheaper for a two-person firm sharing client folders?
Neither, at list price. Claude Team and ChatGPT Business both charge $25 a seat a month billed monthly, or $20 billed annually, so $50 or $40 a month for two people. Claude Team needs at least two seats; ChatGPT Business is listed with a two-seat minimum on OpenAI's pricing page, though its help centre says a new workspace may be able to start with one paid seat (both checked 20 November 2026). Two people need two seats either way. Decide on chat visibility, file limits and memory controls instead of price.
Further reads
- Claude Projects for Executive Assistants: How to Set Them Up — Separate recurring assistant jobs by audience and purpose.
- How to Set Up a Claude Project for Marketing: Voice and Calendar — Organise reusable marketing references and briefs.
- Virtual Assistant Confidentiality Checklist: Using AI Without Leaking Data — Agree client-data handling before involving AI.
- Questions to Ask Before Buying AI That Touches Client Data — Ask vendors about the controls your clients require.
- How to Set Up an AI Assistant for Landscaping Quotes and Replies — Set up a ChatGPT or Claude Project that drafts landscaping quotes and replies from your price list, with ten test enquiries and handover rules.
- How Can Tax Professionals Use Claude and Control Client Data? — One Claude project per job, a citation-check routine and the plan settings that decide what happens to client data.
Show 6 more related tutorials
- Claude or ChatGPT for Lawyers? How to Test Both — A same-task test for a small firm: identical clause summaries and letters in both tools, scored on your own documents.
- What Are the ChatGPT Subscription Options for a Small Team? — The ChatGPT tiers from Free to Enterprise, judged for a small team on seats, admin tools and data controls.
- Claude vs ChatGPT for Startups: Which Plan for a Team of 3 to 10 — Claude Team and ChatGPT Business compared for a three-to-ten-person startup: seat costs, limits, sharing and a one-week trial.
- How to Customise ChatGPT for Business: Projects and Files — Build a ChatGPT Project that knows your tone, price list and policies, then prove it works with ten real customer questions.
- Is Claude Pro Worth It? Pro vs Free for a Small Business — What Claude Pro adds over the free plan, who should pay $17 to $20 a month, and the point where Team makes more sense.
- Why Are My ChatGPT Projects Missing and How Do I Get Them Back? — A checklist for missing ChatGPT Projects in a business: wrong account or workspace, admin and sharing changes, Library becoming Space, and deletions.
Browse all AI-for-business tutorials
Sources: Claude help centre articles on what projects are, creating and managing projects, project visibility and sharing, uploading files, RAG for projects, and chat search and memory; Claude pricing; Anthropic privacy article on model training. OpenAI help articles Projects in ChatGPT and Uploading files and audio to ChatGPT; ChatGPT individual and Business pricing.
