Table of Contents
- Key takeaways
- Can you train ChatGPT on your own data?
- How to choose a method
- Method 1: Upload files in a ChatGPT chat
- Steps
- Cost
- Limits
- Who it fits
- Method 2: ChatGPT projects and custom GPTs
- Steps for a project
- Steps for a custom GPT
- Cost
- Limits
- Who it fits
- Method 3: A no-code chatbot platform trained on your content
- Steps
- Cost
- Limits
- Who it fits
- Method 4: Retrieval (RAG) through the OpenAI API
- Steps
- Cost
- Limits
- Who it fits
- Method 5: Fine-tuning through the OpenAI API
- Steps
- Cost
- Limits
- Who it fits
- Does ChatGPT train on the data you give it?
- Worked example: an online store with 40 help pages
- Common mistakes when training ChatGPT on your own data
- Frequently asked questions
- Basics
- Can you train ChatGPT on your own data?
- How do you train ChatGPT on your own data without code?
- What is the difference between a custom GPT and a ChatGPT project?
- How much data do you need to train ChatGPT?
- Privacy and data
- Does ChatGPT use your data for training?
- Is data sent to the OpenAI API used to train models?
- How long does ChatGPT keep uploaded files?
- Cost
- Can you train ChatGPT on your data for free?
- How much does it cost to train ChatGPT on your own data?
- Is fine-tuning ChatGPT worth the cost?
- Accuracy and updates
- How do you update ChatGPT when your data changes?
- Will ChatGPT stop making things up once it has your data?
- Can ChatGPT write in your style?
- Sources
Created by
Do not index
Created time
Jul 16, 2025 11:07 AM
You cannot retrain ChatGPT itself on your own data, but you can give an OpenAI model your content in five ways: upload files in a chat, use a ChatGPT project or custom GPT, use a no-code chatbot platform, build retrieval through the OpenAI API, or fine-tune a model through the API.
TL;DR
- "Training ChatGPT" almost always means retrieval. The model reads your content at answer time. Its weights do not change.
- For personal use, upload files in a chat or keep them in a ChatGPT project. Projects hold 5 files on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise, and Edu.
- New custom GPTs can only be created in Business, Enterprise, and Edu workspaces. OpenAI plans to retire custom GPTs and recommends Plugins as the replacement.
- For a chatbot that customers use on your website, pick a no-code platform or build retrieval on the OpenAI API. ChatGPT projects and GPTs only work for people signed in to ChatGPT.
- Fine-tuning is no longer a realistic starting point. OpenAI is winding down the fine-tuning platform and new users cannot join it.
Key takeaways
Method | Best for | Cost to start | Main limit |
1. Upload files in a chat | A one-off question about a document | $0 on the Free plan | Free accounts get 3 file uploads a day |
2. ChatGPT projects and custom GPTs | Repeat work inside ChatGPT, alone or with a team | $0 for a project. $20 a month for Plus, billed monthly | Only signed-in ChatGPT users can use it |
3. No-code chatbot platform (SiteGPT) | A customer-facing website chatbot | $59 a month billed monthly, or $39 a month billed yearly | Paid only, with monthly message quotas |
4. Retrieval (RAG) through the OpenAI API | Developers building answers into their own product | Usage-based: $2.50 per 1,000 file search calls plus tokens | Needs code, hosting, and upkeep |
5. Fine-tuning through the API | A fixed output style or format | $100 per training hour on the one model priced today | Closed to new users |
Can you train ChatGPT on your own data?
You cannot train ChatGPT itself on your own data, because the ChatGPT app gives no user a way to change the model's weights. What you can do is give the model your content so it answers from that content.
Three different things get called "training". They are worth separating before picking a method.
What people call training | What actually happens | Does the model change? |
Instructions | You tell the model how to behave: role, tone, rules, format | No |
Retrieval (RAG) | The system finds the relevant parts of your files or pages and passes them to the model with each question | No |
Fine-tuning | You supply example prompts and ideal answers, and OpenAI produces a customized copy of a model | Yes |
Methods 1 to 4 below are all retrieval plus instructions. Only method 5 changes a model. The longer explanation of how retrieval works is in what is RAG.
This guide covers the methods that involve ChatGPT or OpenAI models. For the tool-agnostic version, including how to prepare the content itself, see how to train a chatbot on your own data.
How to choose a method
Choose a method by answering one question first: who will ask the questions. The answer rules out most of the list.
If you need... | Pick... | Why |
An answer from one document, today | Method 1: upload the file in a chat | No setup, and it works on the Free plan |
The same files and instructions in every chat | Method 2: a ChatGPT project | Files and instructions stay attached to the project |
A shared assistant for a team that already pays for ChatGPT Business | Method 2: a shared project or a custom GPT | Workspace members can use it without extra tools |
A chatbot on your website that answers customers | Method 3: a no-code platform | Visitors do not need a ChatGPT account, and no code is needed |
Answers from your content inside your own app | Method 4: retrieval through the API | Full control over the interface, the model, and the data flow |
A fixed tone, classification, or output format | Method 5: fine-tuning, if you already have access | Examples teach style better than instructions do |
Content that changes every week | Method 3 or 4 | Retrieval reads the current content. A fine-tuned model has to be retrained |
Method 1: Upload files in a ChatGPT chat
Best for: a one-off question about a document.
Uploading files in a chat is the fastest way to make ChatGPT answer from your own data. You attach a file to a conversation and ChatGPT reads it for that conversation.
Steps
- Open a new chat in ChatGPT.
- Attach the file: a PDF, document, spreadsheet, or presentation.
- Ask the question and tell ChatGPT to answer only from the file.
- Ask it to quote the passage it used, then check the quote against the file.
Cost
File uploads are available on the Free plan and on paid plans. ChatGPT Plus costs $20 a month, billed monthly, and has higher upload limits. OpenAI does not offer annual billing for Plus.
Limits
- Free accounts can upload 3 files a day. Other users can upload up to 80 files every 3 hours.
- Each file can be up to 512 MB. Text and document files are capped at 2 million tokens each.
- Spreadsheets are capped at about 50 MB and images at 20 MB.
- Outside ChatGPT Enterprise, ChatGPT extracts the text from a document and discards the images inside it.
- The file belongs to that one conversation. A new chat starts without it.
Who it fits
File upload fits one person with one question. It does not fit anything customers or a team need to use repeatedly.
Method 2: ChatGPT projects and custom GPTs
Best for: repeat work inside ChatGPT, alone or with a team.
ChatGPT projects and custom GPTs both keep files and instructions together so every chat starts with the same context. A project is the option open to every plan today.
Steps for a project
- Select New project in the ChatGPT sidebar and name it.
- Add files: PDFs, spreadsheets, documents, images, or pasted text.
- Open Project settings and write the project instructions. State the role, the tone, and the rule to answer only from the project files.
- Optional: paste a link to a Google Drive file or folder, or a Slack channel, as a project source.
- Start a chat inside the project and test it with real questions.
- Select Share to invite other people with chat or edit access.
Steps for a custom GPT
New custom GPTs can only be created by users in ChatGPT Business, Enterprise, and Edu workspaces. Personal accounts on Free, Go, Plus, and Pro cannot create or publish new GPTs. They can still use existing GPTs and edit ones they already made.
- In a Business, Enterprise, or Edu workspace, open Explore GPTs in the sidebar.
- Select Create to open the GPT builder.
- Write the instructions and add a name, description, and conversation starters.
- Upload reference files under Knowledge.
- Test the GPT in Preview, then save it.
OpenAI plans to retire custom GPTs. Its help center lists December 11, 2026 as the planned retirement date for Enterprise workspaces and says other plans are expected to follow the same timeline. OpenAI recommends moving GPT workflows to Plugins, which combine reusable instructions with connected apps.
Cost
Plan | Price | Files per project | Can create new custom GPTs |
Free | $0 | 5 | No |
Plus | $20 a month, billed monthly | 25 | No |
Business, Standard seat | $25 per user a month billed monthly, or $20 billed annually. Minimum 2 seats | 40 | Yes, if workspace settings allow it |
Enterprise | Custom | 40 | Yes, if workspace settings allow it |
A Business workspace needs at least 2 seats. That makes the lowest price for creating a custom GPT $50 a month billed monthly, or $40 a month billed annually ($480 a year).
Limits
- A custom GPT holds up to 20 knowledge files, each up to 512 MB.
- Shared projects allow 5 collaborators on Free, 10 on Go and Plus, and 100 on Pro.
- Only people signed in to ChatGPT can join a shared project. A website visitor cannot use it.
- A Google Drive source added inside a project is searched on demand. It is not synced in advance.
- Chats created with a GPT cannot be moved into a project.
Who it fits
Projects fit individuals and small teams who already work in ChatGPT. Custom GPTs fit teams with a Business, Enterprise, or Edu workspace, and those teams should plan for the move to Plugins.
Method 3: A no-code chatbot platform trained on your content
Best for: a customer-facing website chatbot.
A no-code chatbot platform does the retrieval work for you and gives you a chat widget for your website. You add your content, the platform indexes it, and an OpenAI model answers visitors from that content. Visitors do not need a ChatGPT account.
SiteGPT is the worked example here. Its pricing page names the models it uses: GPT-4.1-mini and GPT-4.1.
Steps
- Start the 7-day trial. It requires a credit card.
- Create a chatbot and add your content. Sources include website URLs and sitemaps, uploaded files such as PDF and DOCX, YouTube, and cloud storage such as Google Drive and Notion.
- Write the instructions: what the bot covers, its tone, and what it should say when the content has no answer.
- Test it with the last 20 to 50 real questions customers asked.
- Copy the embed snippet into your website.
- Turn on escalation so visitors can reach a person.
The full walkthrough with screenshots is in how to create an AI chatbot. The embedding step is covered in how to embed ChatGPT in a website.
Cost
SiteGPT plan | Billed monthly | Billed yearly | Messages a month | Pages of content | Content refresh |
Starter | $59 a month | $39 a month ($468 a year) | 4,000 GPT-4.1-mini or 400 GPT-4.1 | 1,000 | Manual |
Growth | $129 a month | $79 a month ($948 a year) | 10,000 GPT-4.1-mini or 1,000 GPT-4.1 | 10,000 | Monthly |
Scale | $429 a month | $259 a month ($3,108 a year) | 40,000 GPT-4.1-mini or 4,000 GPT-4.1 | 50,000 | Weekly, with a daily scan |
Yearly billing on Starter saves $240 a year against monthly billing ($708 against $468), which is 34%.
Limits
- There is no free plan. The 7-day trial needs a credit card.
- Message quotas are fixed per plan. A GPT-4.1 message uses 10 times the quota of a GPT-4.1-mini message.
- Starter has no chat integrations, no API access, and manual content refresh only.
- Removing SiteGPT branding is a paid add-on at $39 a month billed yearly.
- The pricing page lists two OpenAI models: GPT-4.1-mini and GPT-4.1.
Who it fits
A no-code platform fits a business that wants a website chatbot for customer service and has nobody to build and maintain one. The comparison of AI chatbots for a website covers other platforms in this category. To manage that chatbot from Claude once it is live, follow the steps in building custom AI chatbots with ChatGPT and Claude. If you want a platform with a free plan to test on first, the DocsBot AI review covers its plans and AI credit limits.
Method 4: Retrieval (RAG) through the OpenAI API
Best for: developers building answers into their own product.
Retrieval through the OpenAI API means writing code that stores your files in a vector store and lets a model search them before it answers. OpenAI hosts this as the file search tool in the Responses API.
Steps
- Create an OpenAI API account and an API key.
- Create a vector store and upload your files to it. Supported formats include PDF, DOCX, PPTX, HTML, Markdown, JSON, and plain text.
- Call the Responses API with the
file_searchtool and the vector store ID.
- Read the answer and the file citations that come back with it.
- Build the chat interface, host it, and add a way to refresh the files when content changes.
Cost
The API is billed by usage. File search costs $2.50 per 1,000 tool calls. Vector store storage costs $0.10 per GB per day, and the first GB is free. Model tokens are billed on top at the chosen model's rate.
Model on the API pricing page | Input, per 1M tokens | Output, per 1M tokens |
gpt-6-luna | $0.10 | $0.50 |
gpt-6.1-sol | $2.00 | $10.00 |
gpt-6-astra | $10.00 | $50.00 |
These are standard, short-context prices.
Worked example. Assume 10,000 answered questions a month. Assume each question makes one file search call and uses 3,000 input tokens and 300 output tokens. Storage stays under the free 1 GB.
Cost line | On gpt-6-luna | On gpt-6.1-sol |
File search: 10,000 calls at $2.50 per 1,000 | $25.00 | $25.00 |
Input: 30M tokens | $3.00 | $60.00 |
Output: 3M tokens | $1.50 | $30.00 |
Total a month | $29.50 | $115.00 |
The token counts are assumptions. Real usage depends on how much retrieved text each answer needs. The totals also leave out developer time, hosting, and the chat interface.
Limits
- It needs code. There is no widget, dashboard, or handoff to a person unless you build them.
- File search rate limits start at 100 requests a minute on Tier 1.
- Content refresh is your job. Nothing re-crawls a website for you.
- Abuse monitoring logs can retain prompts and responses for up to 30 days by default.
Who it fits
API retrieval fits a team with developers that wants the answers inside its own product. For a standard website chatbot, the build is more work than a platform. If your code still calls a GPT-3 or GPT-3.5 model, check the shutdown dates and replacement models in what a GPT-3 chatbot means today before you start. The trade-offs are covered in the guide to building a custom AI chatbot.
Method 5: Fine-tuning through the OpenAI API
Best for: a fixed output style or format, for teams that already have access.
Fine-tuning trains a customized copy of an OpenAI model on example prompts and ideal answers. It is the only method here that changes a model.
OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users. Existing users can still create training jobs for now, and fine-tuned models stay available until their base models are deprecated.
Steps
These steps apply to organizations that already have fine-tuning access.
- Set up evaluations first, so you can tell whether the fine-tuned model beats the base model.
- Build a training set of realistic prompts and ideal answers. OpenAI requires at least 10 examples and recommends starting with 50.
- Format the examples as a JSONL file in the chat completions format.
- Upload the file and create a fine-tuning job on a supported base model.
- Evaluate the result against the base model before using it.
Cost
The API pricing page lists one fine-tuning price today.
Model | Training | Input, per 1M tokens | Output, per 1M tokens |
o4-mini-2025-04-16 | $100.00 an hour | $4.00 | $16.00 |
o4-mini-2025-04-16 with data sharing | $100.00 an hour | $2.00 | $8.00 |
OpenAI's supervised fine-tuning guide names gpt-4.1, gpt-4.1-mini, and gpt-4.1-nano as the models to use it with.
Limits
- New users cannot get access.
- Fine-tuning teaches patterns from examples. OpenAI lists classification, nuanced translation, and fixed output formats as its best uses.
- When your content changes, the model does not. You would have to train again.
- A fine-tuned model stops working when its base model is deprecated.
Who it fits
Fine-tuning fits classification, a strict output format, or a consistent style, and only for teams that already use the platform. It does not fit a knowledge base. Use retrieval for facts.
Does ChatGPT train on the data you give it?
ChatGPT may use content from individual accounts to train OpenAI's models unless you opt out. Business products and the API are not used for training by default.
Where your data goes | Used to train OpenAI models by default? | How to change it |
ChatGPT Free, Go, Plus, Pro | Yes, it may be | Turn off Improve the model for everyone under Settings, Data controls |
Temporary chats | No | Nothing to change |
ChatGPT Business, Enterprise, Edu | No | Nothing to change |
OpenAI API | No | Organization owners can opt in to share data |
One detail is easy to miss. If you give a thumbs up or thumbs down on a response, OpenAI may use that whole conversation for training, even after you opt out.
Before uploading company documents to a personal ChatGPT account, turn the training setting off or use a Business workspace.
Worked example: an online store with 40 help pages
This example shows how the choice plays out for one business. The store is hypothetical.
The store has 40 help pages and 3 PDF policy documents. It wants customers to get answers about shipping and returns on its website. Nobody on the team writes code.
Method | Does it work for this store? | Reason |
1. File upload | No | Only the person in the chat can ask questions |
2. Project or custom GPT | No | Customers would need a ChatGPT account, and it does not sit on the store's website |
3. No-code platform | Yes | 40 pages fit inside a 1,000-page plan, and the widget goes on the site without code |
4. API retrieval | Only with a developer | Usage would cost about $29.50 a month at 10,000 questions, but someone has to build and host it |
5. Fine-tuning | No | The store needs current facts, and new users cannot join the platform |
The store picks method 3. Method 2 is still useful internally: the owner keeps the same 3 PDFs in a ChatGPT project to draft policy updates.
Common mistakes when training ChatGPT on your own data
The most common mistakes come from picking a method before deciding who the user is. Five show up often.
- Reaching for fine-tuning to add knowledge. Fine-tuning teaches style and format. Facts belong in retrieval.
- Building a custom GPT for customers. A GPT lives inside ChatGPT, and OpenAI plans to retire the feature.
- Uploading messy files. Scanned PDFs and complex layouts are harder for the model to read. OpenAI recommends clear, text-forward files.
- Skipping the "I do not know" rule. Without it, the model fills gaps from general knowledge. The guide to chatbot hallucinations covers the other guardrails.
- Never refreshing the content. A bot repeats whatever its sources say, including last year's prices.
Content preparation has its own guide in chatbot training.
Frequently asked questions
Basics
Can you train ChatGPT on your own data?
You cannot change the ChatGPT model yourself. You can give it your data through file uploads, projects, custom GPTs, a no-code platform, or the OpenAI API, and it will answer from that data.
How do you train ChatGPT on your own data without code?
Without code, upload files in a chat, add them to a ChatGPT project, or use a no-code chatbot platform. The first two work inside ChatGPT. A platform is the option for a website chatbot.
What is the difference between a custom GPT and a ChatGPT project?
A project is a workspace that keeps chats, files, and instructions together, and every plan can create one. A custom GPT is a packaged assistant that only Business, Enterprise, and Edu workspaces can create now.
How much data do you need to train ChatGPT?
For retrieval, there is no minimum. One document is enough. For fine-tuning, OpenAI requires at least 10 examples and recommends starting with 50.
Privacy and data
Does ChatGPT use your data for training?
On individual plans, OpenAI may use your content for training unless you turn off "Improve the model for everyone" in Data controls. ChatGPT Business, Enterprise, Edu, and the API are not used for training by default.
Is data sent to the OpenAI API used to train models?
No. OpenAI states that data sent to the API has not been used to train its models since March 1, 2023, unless the organization opts in.
How long does ChatGPT keep uploaded files?
Files uploaded in a chat are kept as long as the chat is. After you delete the chat, OpenAI deletes the file within 30 days, with exceptions for security or legal reasons.
Cost
Can you train ChatGPT on your data for free?
Yes, in a limited way. The Free plan allows 3 file uploads a day and projects with up to 5 files. A customer-facing chatbot costs money through either a platform or the API.
How much does it cost to train ChatGPT on your own data?
It costs $0 for a Free-plan project, $20 a month for Plus, and at least $50 a month billed monthly for a ChatGPT Business workspace. SiteGPT starts at $59 a month billed monthly. API retrieval costs about $29.50 a month for 10,000 questions on the cheapest listed model.
Is fine-tuning ChatGPT worth the cost?
For most teams, no. New users cannot access OpenAI's fine-tuning platform, and the one listed price is $100 per training hour. Retrieval covers most use cases at a lower cost.
Accuracy and updates
How do you update ChatGPT when your data changes?
In a project or GPT, replace the files. On a platform, re-sync the content. SiteGPT refreshes monthly on Growth and weekly on Scale, and Starter is manual. With the API, re-upload the changed files to the vector store.
Will ChatGPT stop making things up once it has your data?
No. Retrieval lowers the rate of invented answers. It does not remove them. Tell the model to answer only from your content, and test it with questions your content does not cover.
Can ChatGPT write in your style?
Yes. Paste 3 to 5 samples of your writing into project instructions and ask ChatGPT to match them. Fine-tuning can also teach style, but new users cannot access it.
Sources
All sources were checked on October 4, 2026.
Written by
