To create a custom GPT, you open the GPT builder in ChatGPT, write the instructions, add knowledge files, switch on capabilities, test it and set who can use it. You can no longer do that on every ChatGPT account, though, and OpenAI is going to phase out custom GPTs altogether. If you build one today, you are building on a feature with an end date. Below you will find what changes and when, what to do with your existing GPTs, how to create a custom GPT while you still can, and when a custom GPT is no longer a good choice.
What is happening to custom GPTs, and when?
OpenAI is phasing out custom GPTs in stages, not on a single date. Its own help pages say that OpenAI plans to retire custom GPTs and move users to Plugins, which bring together reusable instructions and connected apps.
The first change has already happened and concerns building GPTs. Creating and publishing new GPTs is no longer possible on personal ChatGPT accounts, which covers Free, Go, Plus and Pro. It is only still possible in a Business, Enterprise or Edu workspace, provided the workspace settings and permissions allow it. According to OpenAI, timing and availability may differ by plan and workspace, so check what is still possible in your own workspace.
The second is the migration. With the Migrate to plugin option in My GPTs, you can move a GPT to a plugin as soon as that option is available for your account or workspace. That does not happen everywhere at the same time.
The third is the retirement itself. OpenAI has scheduled the retirement on all plans for 11 December 2026. Only Enterprise workspaces with an approved extension keep their GPTs until 11 February 2027. Existing GPTs keep working until then, and you can continue to use and edit them as long as your plan and permissions allow.
OpenAI sets out in its own documentation what does and does not carry over in the migration. A GPT’s instructions become a skill in the new plugin and the knowledge files become reference files, but custom actions and the chosen model do not carry over. OpenAI says nothing about the conversation starters. So do not assume everything will come across automatically, and keep your own copies of your instructions and source files outside ChatGPT.
What is a custom GPT?
A custom GPT is a tailored version of ChatGPT that works according to preset instructions and can optionally access your own knowledge files. Where standard ChatGPT needs context again in every conversation, a custom GPT already has its task, its writing style and the sources it may consult set in advance.
The idea is simple. Instead of typing the same long prompt every time, you configure that prompt once in a custom GPT. The next time you do the same task, you open the GPT and only give the specific input for that round. The rest is already set up.
Under the bonnet, a custom GPT runs on the regular ChatGPT model. So you do not train the model, and you do not add new knowledge to the language model itself. What you do is steer the model’s output through instructions and optional knowledge files.
Who can still create a custom GPT?
You can now only create a custom GPT within a Business, Enterprise or Edu workspace, and only if the admin has allowed it in the workspace settings. On Free, Go, Plus and Pro, creating and publishing new GPTs is no longer available.
The difference is bigger than it looks. Until recently, a Plus subscription was the normal route to building your own GPT, and almost every guide online is based on that. If you now follow a tutorial on a Plus account that promises your first GPT within half an hour, you will not get past the first step.
Existing GPTs are a different story. They keep working until the retirement, and you can still edit them as long as your plan and permissions allow. So a GPT you built on Plus last year will not disappear today, but no new ones will be added.
What is more, creating and editing only works through the web interface. The mobile apps support using GPTs, but not building them.
What to do now with your existing custom GPTs
Your existing custom GPTs call for three actions in the short term, and you can do the first two today without waiting for the migration environment. The aim is for your way of working to depend not on ChatGPT but on documents you manage yourself.
Take stock of which GPTs you actually use
First take stock of which GPTs you actually use, because in practice that is fewer than your list suggests. Go through your list and note for each GPT when you last opened it, what task it does and who else uses it besides you.
Do not migrate anything you have not opened in six months. A retirement is a good moment to cut back instead of dragging everything along to a new environment. What remains is usually a handful of GPTs that do a real task in your work.
Keep instructions and source files outside ChatGPT
Store the instructions and source files of every GPT you want to keep outside ChatGPT, in your own documents folder or repository. Copy the full instructions field into a text file per GPT, and keep the uploaded source files in their original form.
The reason is simple. Because custom actions and the model choice do not carry over in the migration, and OpenAI does not describe every component, the only certainty is a copy you manage yourself. An instructions field you have polished over six months is work you do not want to redo because a migration tool skipped a field.
Record straight away why each instruction is there. A rule like ‘never mention search volumes’ will be baffling in three months without the reason behind it, and it is precisely those reasons that decide whether you need it again in a new environment.
Prepare for the move to Plugins
Prepare for the move to Plugins by deciding for each GPT what it needs in the new environment. OpenAI describes Plugins as the place where reusable instructions and connected apps come together, and advises you to move your workflows there.
To do that, split each GPT into three parts. The behaviour: what the GPT must always do. The sources: what it may draw on. And the connections: the external systems it calls. That third category is the most fragile, because a connection often has to be set up again in a new environment.
Hold off on moving anything until the migration environment is available in your workspace. Until then your GPT keeps working, and moving everything ahead of time by hand only doubles the work.
When does a custom GPT still make sense, and when not?
A custom GPT still makes sense if you work in a Business, Enterprise or Edu workspace, have a task that comes back often, and have a fixed way of working you want to apply consistently. Examples are drawing up briefings, writing messages in a fixed style, handling emails in a fixed format, or running a type of analysis with the same steps every time.
When I used ChatGPT intensively, I built several custom GPTs for my SEO work: one for keyword research, one for search intent classification, one for competitor analysis, one for content briefs, one for meta titles and meta descriptions, and one for detecting cannibalisation between pages. All those tasks came up often in that work, and that is exactly where the build time paid off.
A custom GPT does not make sense for one-off tasks, for work where every assignment is completely different, or for tasks where the output can no longer be checked afterwards. For exploratory questions, sparring sessions or brainstorms, an ordinary ChatGPT conversation works better, because a custom GPT keeps to its brief rather than ranging freely.
With the retirement in sight, there is another reason not to. Building a new GPT for a way of working you want to use for years is work you will soon have to set up again. For something you need until December, that is fine. For anything structural, you are better off going straight to the environment that will remain.
Confidential business information that you do not want others to be able to retrieve does not belong in a custom GPT either. Depending on the settings and file type, uploaded source files may be readable by whoever uses the GPT. If you share a GPT more widely, bear in mind that its sources may also end up in users’ hands.
What do you need to create a custom GPT?
To create a custom GPT you need three things: access to a Business, Enterprise or Edu workspace where building is allowed, a clearly defined task, and source material or examples to steer the output.
The preparation matters more than many people think. Before you open the builder, write down briefly what exactly the GPT should do, for what type of input, with what type of output and in what style. If you skip this step and go straight into the builder, you will find yourself halfway through still working out what the GPT is supposed to do.
That written version is now doubly useful. It is the basis for your instructions field, and it is also the document you will later take with you to Plugins.
How to create a custom GPT in nine steps
You create a custom GPT in ChatGPT’s GPT builder, in nine steps. Below they are grouped, with notes for each group on what to watch for and where most people get stuck.
Steps 1 to 3: from goal to the right build mode
Define the goal and scope of your custom GPT before you type anything into the builder. Write on one page what the input is, what the output should be, who it is for, and what the GPT explicitly does not do. That last question is the most important, because a GPT that tries to do everything does nothing well.
Then open the GPT builder from the sidebar. You arrive in an environment with the build functions on the left and a preview window on the right, where you can test the GPT straight away. That preview is not an afterthought. With every change to the instructions or sources, you can see immediately whether the effect is right. If you only test at the end, you miss the small deviations that are still easy to correct at this stage.
Next, choose between the conversational builder and the configure view. In the conversational builder, ChatGPT asks you questions and fills in the fields based on your answers. In the configure view, you fill in everything yourself. The configure view works better for most people, because the conversational builder produces instructions that are longer than necessary and that you end up rewriting by hand anyway. So start straight away in the configure view, where name, description, instructions, conversation starters, knowledge and capabilities are neatly listed one below the other.
Steps 4 and 5: instructions and knowledge files
Write your instructions as you would a briefing document for a new employee. The instructions field is the most important part of the whole GPT and determines how it behaves in every conversation, so treat it as a fixed system prompt that is always active.
Good instructions include a short role description, a defined list of tasks, the steps the GPT goes through before it answers, the output format, and clear limits on what it does and does not do. For multi-step workflows, use an explicit step structure, separate sections clearly, and give specific instructions rather than a long list of prohibitions. Examples of acceptable and unacceptable output set the standard in a way a description cannot.
Then upload your source files in the knowledge field. You can add up to 20 files per GPT, with a maximum of 512 MB per file. In practice, many small, focused files work better than one huge file the GPT has to search through itself, even though those limits sound generous.
Text files without complex formatting are preferable. PDFs with tables, columns or lots of images produce worse answers than a simple Markdown or text file with the same content. The cause is not the upload function but how a model reads documents: a clean textual structure is always easier to process than a visual design.
Steps 6 and 7: switching on capabilities and testing systematically
Switch on only the capabilities the GPT really needs. The standard options are web search, image generation and the code interpreter for data analysis. You can also set up actions that let the GPT call external APIs.
Only switch on web search if the GPT needs current information. For a GPT that works with your own source files it actually gets in the way, because it then searches externally instead of sticking to your own documents. The code interpreter is useful for data, spreadsheets and calculations, image generation only for visual output, and actions are an advanced option that falls outside the scope of a first custom GPT.
Then test systematically in preview mode rather than ad hoc. Prepare five to ten test cases in advance that cover the full breadth of your GPT: a typical assignment, an edge case, an assignment where the GPT should refuse or ask for more information, and an assignment where the input is incomplete.
What you test is not only whether the output is correct, but also whether the GPT sticks to the behaviour and tone rules in your instructions. Deliberately ask for things that are out of scope and see whether it recognises them. With every deviation, go back to the instructions or the sources, not to a new test round with the same settings, because a deviation repeats itself until you tackle the cause in the configuration.
Steps 8 and 9: visibility and maintenance
Set your GPT’s visibility in the publishing settings, and choose deliberately. For business use, sharing internally within the workspace is usually the right choice for GPTs that contain the organisation’s own content or ways of working. Reserve wider sharing for GPTs without confidential information, because in practice your instructions and source files can then be retrieved by anyone who makes the effort.
Maintenance starts the moment the GPT goes live. In practice, instructions sometimes work out differently from what you expected, sources go out of date and your own way of working evolves. ChatGPT keeps a version history, so you can restore an earlier version if a change makes things worse. Testing small changes and rolling them back if needed is safer than a major overhaul you can no longer compare with the original.
From now on, factor the retirement into that maintenance too. Every time you update a GPT, also update the copy of the instructions outside ChatGPT, so that when you move to Plugins you take the current version with you and not one from months ago.
The difference between instructions and knowledge files
The difference between instructions and knowledge files is often underestimated, yet it is the most important configuration decision for a custom GPT. Instructions determine how the GPT behaves, knowledge files determine what it can draw on.
In the instructions you put the role, tasks, way of working, output format, style guidelines and limits on what the GPT does and does not do, and all of that is behaviour. In the knowledge files you put documentation, sample texts, product information, glossaries and policy documents, and all of that is source material.
A common mistake is pasting entire policy documents into the instructions, or uploading behaviour rules as a source file. Both work worse than they should. Policy documents in the instructions make those instructions so long that the GPT skips important parts. Behaviour rules as a source file are treated as reference material rather than as an obligation.
The rule of thumb: everything you want the GPT to do every time belongs in the instructions, and everything it may optionally draw on when a question calls for it belongs in the knowledge files. You will take that same separation with you to Plugins, where reusable instructions and connected sources are again two different things.
Which model runs under a custom GPT?
The model that runs under a custom GPT has changed several times in 2026, which explains why a GPT may behave differently from when you built it. On 13 February 2026, GPT-4o, GPT-4.1, GPT-4.1 mini and the original GPT-5 in its Instant and Thinking variants were removed from ChatGPT. The GPT-5.1 models have no longer been available since 11 March 2026.
As a result, a custom GPT runs on ChatGPT’s current line-up, and when building it you can set a recommended model that users are offered when they start a new conversation. GPT-5.6 Sol rolled out to eligible paid plans from 9 July 2026, and on 6 August 2026 GPT-5.6 Luna became the default model for Free and Go. GPT-6 Astra arrived in early September 2026. In ChatGPT, Astra runs in Work and Codex, and in the chat only as GPT-6 Pro on Pro, Business and Enterprise (not on Plus). The model is also available through the API, Microsoft Azure and AWS Bedrock. Since 7 October 2026, Chat has run on GPT-6: Plus, Pro, Business and Enterprise use GPT-6 Sol by default, and Free and Go have used GPT-6 Luna since 8 October. OpenAI o4-mini was removed from ChatGPT on 13 February 2026 and o3 on 26 August 2026.
What this means for you: a custom GPT you built a year ago now runs on a different model, possibly with different behaviour. Instructions that worked well on the old model may turn out slightly differently on the new one. With every model change, it pays to run your most important GPTs through your standard test cases again and adjust the instructions where needed.
A new model version is not automatically a better model, and I see that with Claude too, which I use every day. When I moved from Opus 4.6 to 4.7, I noticed that the answers became longer and less sharp, while the model was more capable in other respects. The same principle applies to ChatGPT: a new model is not automatically a better model for your specific prompt.
The same goes for new features. ChatGPT Work, the agent for longer tasks on paid plans that has been available since 9 July 2026 and replaced ChatGPT agent mode, carries out tasks on its own and therefore calls for different instructions from those for a custom GPT.
Common mistakes when building a custom GPT
Common mistakes when building a custom GPT are rarely technical. They lie in the scope, the instructions and the way output is checked. Below are the mistakes I came across consistently during the time I worked intensively with ChatGPT.
The first mistake is a task description that is too broad. ‘A GPT for my marketing work’ produces a GPT that does everything half-well. ‘A GPT that writes messages in this format, in this style, with this maximum length’ produces a GPT that does one thing well. Starting narrow works better than trying to go broad.
The second mistake is instructions phrased as wishes rather than rules. ‘Try to write in a professional tone’ is a wish. ‘Write in plain business language, avoid jargon, use short sentences’ is a rule. In production, wishes are ignored more often than rules.
The third mistake is source files with visually complex formatting. PDFs with multi-column layouts, tables that run across pages or tables of contents that refer to page numbers produce worse answers than the same content in plain text. If you convert a PDF to Markdown before uploading, you will see the output improve noticeably.
The fourth mistake is ignoring the length of the instructions field. With longer instructions containing several conditions, a GPT follows some of the instructions and ignores others without saying so. For more complex tasks it therefore pays to split up the instructions, or to carry out the task in steps across several prompts, instead of giving one long instruction that is partly skipped in production.
The fifth mistake, and in my experience the most dangerous, is assuming the GPT is right without checking. One specific pitfall I consistently saw with my own GPTs for SEO work is that ChatGPT adds a search volume to a keyword unprompted, even though there is no connection to Ahrefs or any other keyword tool. Those figures are hallucinations that sound plausible to anyone without access to keyword data. I saw the same pattern with made-up sources, non-existent Python functions and summaries that sound plausible but differ in substance from the original.
A custom GPT that produces wrong output is no safer than an ordinary ChatGPT prompt. It merely produces that wrong output more consistently. That is why a checking step remains indispensable, even with a GPT you built yourself. From my years in the banking sector I know that principle as the four-eyes principle: nothing goes out the door without a second check. With AI output, that second check matters even more than with human work, because AI errors are phrased with more confidence.
Custom GPT versus the alternatives within ChatGPT
A custom GPT is not the only way to preconfigure behaviour within ChatGPT, and with the retirement in sight, that comparison is all the more relevant. ChatGPT offers custom instructions, projects, custom GPTs and Plugins for this.
Custom instructions are account-wide preferences. You fill in once who you are and how you want ChatGPT to communicate with you, and those settings apply to all your conversations. They suit universal preferences such as language, tone and level of knowledge, but not specific tasks, because you only have one set of custom instructions per account.
Projects are workspaces where you group related conversations, with their own instructions and optional files that apply within that project. They suit long-running jobs where several conversations belong together, such as a research project or a quotation process. If you can no longer create a custom GPT, creating a ChatGPT project is in many cases the closest alternative that works today.
Custom GPTs are reusable configurations with instructions, sources and optional capabilities that you can share and publish. They suit recurring tasks you want carried out consistently, possibly by others too. Creating them is only still possible in a business workspace, and the feature is being phased out.
Plugins are the route OpenAI is pointing to, bringing together reusable instructions and connected apps. Since 9 July 2026, Plugins have been listed in the Plugin Directory, which replaced the App Directory, and OpenAI describes them as bundles of instructions and skills with connected apps (connectors). The direction is therefore settled: your workflows belong there.
A custom GPT does not keep personal context between users, whereas ChatGPT’s memory does hold on to information about you.
So the rule of thumb remains: custom instructions for how you want ChatGPT to talk to you in general, projects for a defined job with several conversations, and a reusable configuration for work that comes back regularly. That reusable configuration is a custom GPT for as long as that is possible, and a Plugin after that. Building your own configuration is fine for an individual workflow, but it only becomes an organisation-wide tool once the instructions, maintenance and version control are set down for the whole team.
Custom GPT versus Claude Projects and Claude Skills
Comparing a custom GPT with Claude Projects has suddenly become topical because of the retirement, since many professionals work with Claude as well as ChatGPT. Both let you set up a workspace with fixed instructions and source files, with some relevant differences.
A Claude Project works on the basis of a system prompt you write yourself and source files you add to the project. The conversations you have in that project automatically have access to those instructions and files. The configuration is deliberately simple: no capability toggles, no build mode, no store. You write instructions, upload files and get to work.
In my daily work I use Claude Projects intensively, and before that I built several custom GPTs. What struck me is that those custom GPTs felt more like a watered-down version: the same basic function, but with less consistent results on comparable tasks. That observation is mine and not a measurement, and if you use ChatGPT as your main tool and set up the instructions and sources well, you can get a long way with it.
Anthropic also offers Skills, a separate feature that is more like reusable behaviour modules than a workspace. Skills, Projects and a custom GPT each solve a different problem: skills for repeatable behaviour patterns, projects for defined jobs, and a custom GPT for a reusable configuration with a wider audience.
If you do not use Claude yet, you do not need to switch to work effectively with AI. If you use both platforms, you would do well to decide for each type of work where it belongs, instead of reaching for one platform by default.
Frequently asked questions about custom GPTs
The questions below about custom GPTs cover the retirement, the plans on which building is still possible, what happens to existing GPTs and which alternative comes closest.
Can I still create a new custom GPT on ChatGPT Plus?
You can no longer create a new custom GPT on ChatGPT Plus. Creating and publishing new GPTs is not available on personal accounts, including Free, Go, Plus and Pro. It is only still possible in a Business, Enterprise or Edu workspace, and then only if the workspace settings and permissions allow it.
When exactly will custom GPTs disappear?
According to OpenAI, custom GPTs will disappear on 11 December 2026, on all plans. Only Enterprise workspaces with an approved extension keep them until 11 February 2027, and existing GPTs keep working until the retirement.
What happens to my existing custom GPTs?
Your existing custom GPTs keep working until the retirement, and you can continue to use and edit them as long as your plan and permissions allow. With Migrate to plugin in My GPTs, you can move a GPT to a plugin as soon as that option is available for your account. Because custom actions and the model choice do not carry over, store the instructions and source files outside ChatGPT now.
Is a ChatGPT project a good alternative to a custom GPT?
For many tasks, a ChatGPT project is the closest alternative, because there too you set instructions and files that apply to all conversations within that project. The difference lies in sharing and publishing: a project is your workspace, not a tool you offer to others. For a way of working that only you use, that makes little difference. For a tool for your team, it does.
Can I create a custom GPT without a paid subscription?
You cannot create a custom GPT without access to a business workspace. Free and Go users can open and use existing GPTs made by others, but building your own is only possible in a Business, Enterprise or Edu workspace where it is allowed, and only through the web interface, not the mobile app.
Where can I learn how to set this up in my own work?
Learning how to set this up in your own work goes beyond clicking through a builder. It is about defining the task, writing instructions that are actually followed in production, and building in a checking step that catches errors before they go any further. LearnLLM’s ChatGPT Course covers this within a broader way of working, so that a change of platform does not upend your way of working.
Building your own configuration is not a technical project but a work project. The effort you put into scope, instructions and testing decides whether it will save you time or amplify errors, and that effort moves with you when the platform changes. If you want to build that way of working with the checkpoints included, sign up for the ChatGPT Course and build a way of working you can account for to clients and colleagues.
