All three save you from re-explaining yourself to an AI. A Custom GPT is a saved ChatGPT setup that OpenAI is retiring, a Claude Project is a workspace with its own files, instructions and memory, and an agent is a separate system that remembers you, works on a schedule and can act within limits you set.

The products moved a lot in 2026. OpenAI has stopped new GPT creation on personal accounts and plans to retire custom GPTs in favour of plugins, with 11 December 2026 as the standard retirement date (OpenAI help centre). Anthropic has added memory to Projects, offers scheduled tasks on paid plans, and is testing a new version of Projects built around work that carries on after you close your laptop. Choose for the next twelve months, not the next five years.

Disclosure: BeLateral makes Able°, which gives you your own AI agent, so weigh our view accordingly.

How do the three compare?

Custom GPTs Claude Projects AI agent
Lives in ChatGPT Claude Its own system, usually reached through a chat app
Context Instructions plus up to 20 knowledge files Project knowledge, instructions and project memory (when memory is on) Persistent memory across conversations
Runs on a schedule No On paid plans, through scheduled tasks Yes
Takes actions Limited, through configured actions or apps On paid plans, through Cowork connectors and plugins Yes, within permissions
Shared with a team Yes, by link or workspace Yes, on Team and Enterprise plans Yes, with roles
Cost Inside a ChatGPT Business plan ($25/user/month, $20 billed yearly); no new creation on personal plans Free for up to 5 projects; Pro $20/month ($17 yearly); Team $25/seat ($20 yearly) Setup, hosting and model usage, depending on the provider
Best for A repeatable assistant task, until retirement Research and document-heavy work Ongoing work that happens without you

Plans and prices checked 24 October 2026 against Anthropic's pricing page and OpenAI's Business plan help article. Prices move, so confirm them before you budget.

What is a Custom GPT?

A Custom GPT is a saved configuration of ChatGPT. You write instructions, attach reference files, optionally add actions that let it call another service, and share it by link or inside a workspace. OpenAI's help centre lists the parts: name, description, conversation starters, instructions, knowledge and capabilities.

Where it works well. Packaging expertise. A good GPT for proposal writing, or for questions about one specific policy, spares a team from setting up the same context every time someone opens a chat. The link is the product: one address, one behaviour, no training needed.

Where it falls short. Each session still starts without the history of your work. ChatGPT's own memory may carry some facts between chats, depending on your settings, but a GPT has no separate memory: it knows its instructions and files, not what you decided last month. Nothing happens while you're away, either. OpenAI's scheduled tasks don't run inside custom GPTs.

The bigger issue in 2026 is retirement. OpenAI's help centre says new GPT creation and publishing are unavailable on personal accounts (Free, Go, Plus and Pro), while Business, Enterprise and Edu workspaces can still create them where permissions allow. The replacement is plugins, and OpenAI's retirement and migration FAQ sets out what survives the move.

When to keep one. If your ChatGPT Business workspace has GPTs in daily use, they work today. Inventory them before you replace anything. Migrating a good GPT takes an afternoon, while rebuilding the team habit around it takes a month.

What is a Claude Project?

A Project is a workspace in Claude with its own chat history and knowledge base. You upload documents and write project instructions, and every conversation inside the project draws on both.

Where it works well. Long documents and stable reference material. If a set of files should inform every answer, a Project stops you pasting them again. Paid plans add retrieval, which Anthropic says expands project knowledge capacity by up to ten times as the material nears the context limit, useful when your reference set is a shelf rather than a folder. Project instructions also keep a team's writing consistent inside that workspace.

Where it falls short. On the free plan, a Project is a conversation with better context: Claude answers, you act. Paid plans close much of that gap. Each project has its own memory, on by default for Free, Pro and Max and controlled by owners on Team and Enterprise. Scheduled tasks are available on all paid plans, a Cowork project can hold its own, and Anthropic says they run remotely, even when your computer is asleep. The boundary is that everything happens inside Claude, through the connectors and plugins you enable there.

Free accounts can create up to five projects, so trying one costs nothing. A new version of Projects is also in staged beta, starting with Claude Code on Pro and Max. There, a project is a single conversation that splits work into parallel threads running in the cloud, and Anthropic says the threads continue if you close your laptop. Projects moved to the new version keep their instructions, files, chat history and scheduled tasks.

Treat the new version as a direction, since access is staged and there's a waitlist. Memory and scheduled tasks are available now. Details are in Anthropic's help centre: What are projects? and Schedule recurring tasks in Claude Cowork.

What is an AI agent?

An agent holds a goal, uses tools, remembers across conversations and can run a task at a set time. Its state lives outside any single conversation, and it has a clock.

Where it works well. Continuity. It remembers how you work, messages you without being asked, and carries out a procedure rather than describing one. Many agents can save a procedure as a reusable skill, so a format you settled once never needs explaining again. A Custom GPT helps when you sit down to write; an agent helps when you haven't sat down at all.

Where it falls short. Setup, cost and trust. It needs configuration, examples and review, and it costs more than a chat subscription because it includes infrastructure. Trust comes slower too, because a wrong action matters more than a wrong answer and may have happened before you read the message.

It also only knows what reaches it. If nobody has pasted the tender, forwarded the email or mentioned the deadline, the agent doesn't know about it.

What does each one store?

This is what decides most purchases, and what product pages rarely spell out.

A Custom GPT stores a configuration. Instructions, up to twenty knowledge files of up to 512 MB each according to OpenAI's documentation, and the definition of any actions you connected. It keeps nothing you concluded together.

A Claude Project stores a workspace and, with memory on, a memory. Project instructions, project knowledge, the conversations inside the project and a project memory Claude builds as you work. What it learns in one project stays out of your others, though by default it also uses your account memory. Mention that a client hates unannounced calls and it can carry that forward. The files and instructions are still yours to maintain.

An agent stores a working relationship. Facts about your work, preferences, corrections, decisions and the procedures it has learned. It gets more useful over time and can also drift, because a stored fact has no expiry date. The habits that keep it accurate are the ones that keep any notes accurate: say when something changes, and check what it thinks it knows.

So a GPT is a document you maintain, while an agent (and increasingly a Project with memory on) is a record you correct. Both take effort, and the record is the one that can go stale without anyone noticing.

Which is better for each job?

Context that persists. The agent, and a Claude Project with memory on: both carry forward what you said last month. A GPT keeps only what's in its instructions and files.

Knowledge of your documents. Projects and GPTs, because uploading reference material is how they're designed to work. Agents can use documents too, usually when you supply them in the conversation.

Acting on your behalf. The agent, which is built for it. Claude on a paid plan can act through Cowork connectors and plugins, and a GPT can call a configured action. In every case, what it can do depends on what you connect.

Working while you're away. The agent, or Claude on a paid plan with scheduled tasks. A Custom GPT waits for you to open it.

Ease of setup. Projects and GPTs, comfortably. Either can be useful within an hour.

Team sharing. All three. Projects add view and edit permissions on Team and Enterprise; agents add roles and an activity log, which matters once more than a few people share one.

Cost. Chat subscriptions are cheaper and predictable. An agent carries setup cost plus usage, and usage is the part people underestimate.

What happens to what you've already built?

From a Custom GPT to a plugin. OpenAI's migration FAQ has the detail. Instructions become a skill inside the new plugin, knowledge files are copied into reference files, and connected apps are added as apps. The selected model, custom actions and sharing settings don't carry over. A migrated plugin starts private, and existing users aren't given access automatically. Previous conversations stay with the original GPT, and drafts and unpublished edits aren't included, so publish anything you need first. You can only migrate GPTs you created yourself, so find the original builder of any GPT your team relies on.

For affected Enterprise workspaces, OpenAI's timeline puts the end of new GPT creation at 26 October 2026, standard retirement at 11 December 2026, and 11 February 2027 for workspaces with an approved deferral. Other plans are expected to follow. Check the notice on your own workspace rather than trusting a date in an article, including this one.

From a Custom GPT or Project to an agent. What carries over is the thinking. Your instructions are a prompt you can paste, and your knowledge files are documents you keep. The wiring doesn't transfer, and there's usually less of it than people fear. A team with one GPT and twelve reference files has an afternoon of work; a team with twenty GPTs and custom actions has a small project, and should start with the three people use most.

The habit. People stop using a tool when the place they used it disappears, so a migration done silently in the background tends to end with an empty workspace six weeks later. Tell the team what replaced what, and keep the same name if you can.

A survey firm's year with a GPT

A five-person building survey firm bids for work. One director built a Custom GPT with the firm's tone rules and twelve past proposals, and for a year it earned its keep: draft sections came back in the house style rather than sounding generic.

Then the cracks showed. Tenders arrived at odd hours, nobody wanted to open a chat at nine at night to start one, and the GPT couldn't begin anything on its own. The director who built it moved to another role, nobody knew which of the twelve files were current, and the GPT kept quoting a fee structure from two years earlier. When the firm compared a draft with a recent winning bid, the gap was context the GPT never had: who the client was, what they'd said on the phone, what the firm had promised last time.

They tried a Project next. It suited the long reference documents better, and the instructions were easier to keep tidy, but on the free plan it still waited for someone to open it.

What they needed was a message on Monday morning listing the week's deadlines, a first draft assembled from what they pasted in, and a memory of how each client likes to be written to. That's a schedule, a skill and a memory: the definition of an agent, and by now much of what a paid Claude plan offers too. The deciding questions were where the team already worked and where they wanted the Monday message to land.

The agent still didn't know a tender existed until someone forwarded it. Its value was in assembling the draft and sending the reminder, while finding the work stayed with the team.

What do people get wrong?

Expecting memory from a Custom GPT. It won't remember your last conversation, only what's in its instructions and knowledge.

Using a Project as the team's knowledge base. Fine until the files go stale and nobody owns them. Then it cites last year's policy with total confidence.

Treating an agent as a smarter GPT. The model may be the same. What differs is engineering: an agent persists, schedules and acts.

Ignoring the retirement notice. If a workflow depends on a Custom GPT, put the retirement date in the calendar now, with an owner and an internal deadline a few weeks earlier.

Who should choose which?

Keep a Custom GPT for now if you're in a ChatGPT Business, Enterprise or Edu workspace, the GPT already works, and the job is repeatable drafting or answering questions. Migrate on your own timetable, well before the date in your notice.

Choose a Claude Project if your work is document-heavy, you need a stable reference set and you already use Claude. It's the cheapest good option here: free accounts can try it, and paid plans add scheduled tasks.

Choose an agent if the work is continuous, happens at awkward times, or needs action rather than advice. Read how agent memory works first, because memory is the feature you'll spend the most time correcting.

Choose two if you want a cheap assistant for drafting and an agent for recurring operational work. Plenty of people run exactly that pair.

Wait if you can't yet name three tasks you'd hand over. Paying more won't make an undecided tool useful. AI usage vs an agent is a good place to start.

Common questions

Will my Custom GPT conversations disappear?

OpenAI's migration FAQ says previous conversations stay attached to the original GPT and don't move to the new plugin. Check the retirement notice in your workspace for what happens to access after the retirement date, and copy out anything you'll need to refer to later.

Do Claude Projects remember previous conversations?

Yes, with memory on. It's on by default for Free, Pro and Max, and on Team and Enterprise an owner decides. Each project keeps its own memory, used alongside your account memory but never another project's, and paid plans can also search past chats within a project.

Can I move a Custom GPT into an agent?

You can reuse the instructions and documents, which is most of the value. The configuration doesn't transfer automatically, so expect to rewrite the instructions for the new system and test them with three real prompts plus one hard case before relying on it.

Is an agent more accurate?

Not necessarily. It has more context and can act, which makes it more useful and more consequential when it's wrong. Accuracy depends on the material you give it and the corrections you make, whichever kind of tool you use.

Can a Claude Project run something on a schedule?

Yes, on paid plans. Scheduled tasks work on Pro, Max, Team and Enterprise in Claude Cowork and the new Claude experience, a Cowork project can hold its own, and Anthropic says they keep running when your computer is asleep. On Team and Enterprise, admins control Cowork access.