How to Import Your Claude Agents into ChatGPT: The 2026 Migration Guide for UK Teams

To import your Claude agents into ChatGPT, open the ChatGPT desktop app, go to Settings, Import, select Claude Code or Claude Cowork as the source, choose the instructions, settings, skills, plugins, projects and recent work you want to bring over, then confirm. The import takes minutes, runs one way, and leaves your Claude account exactly as it was. Codex CLI users type /import instead. What does not transfer automatically is claude.ai chat Projects, which you rebuild by hand.

By Toni Dos Santos, Co-Founder, Spicy Advisory. Published 13 August 2026. Written for UK teams who built real work into Claude and are now being asked to justify the invoice.

Why UK teams are moving from Claude to ChatGPT in 2026

Almost nobody switches AI vendors because of a benchmark. They switch because a finance business partner asked a question they could not answer.

The pattern we see across UK mid-market clients is consistent. A team starts on Claude Pro seats bought on a company card. The work is good, so it spreads. Someone adds Claude Cowork for agentic tasks. Someone else runs Claude Code. Twelve months later the AI line on the P&L has three vendors on it, two of them billed in dollars, and one of them is billed partly on usage that nobody can predict a month ahead.

Anthropic changed the shape of that bill in 2026. Bundled token allowances came out of the Enterprise seat, the seat price dropped toward a flat $20, and usage moved onto its own line at API rates. That is more honest pricing, and for light users it is genuinely cheaper. For a team running long-context analysis or agentic loops all day, it turns a fixed cost into a variable one. Variable costs are what get escalated.

OpenAI sells the opposite shape. ChatGPT Business is $25 per user per month billed monthly, or $20 billed annually, with usage inside the number. ChatGPT Enterprise has no published price, and 2026 procurement reports converge on $45 to $75 per seat per month with a reported seat minimum and annual prepay. You pay more per head at the top tier, and in exchange the number stops moving.

"Finance teams do not hate spending money on AI. They hate a line item that changes 40% month to month with no explanation. Half the migrations I have run in the last year were solved by better cost visibility, not by changing vendor. The other half genuinely needed to move."

Toni Dos Santos, Co-Founder, Spicy Advisory

Before you migrate anything, it is worth putting a number on what you are actually spending and where it goes. We wrote a full method for that in AI spend visibility across Claude, ChatGPT, Copilot and Gemini, and there is a shorter list of usage controls in 18 tactics to stop burning Claude credits at work. If the answer after that exercise is still "move", the rest of this guide is the how.

If you are still deciding rather than executing, our head-to-head comparison in Claude vs ChatGPT for business in 2026 covers the capability side, and ChatGPT Enterprise vs Copilot vs Gemini covers the three-way enterprise decision.

What actually transfers, and what does not

This is the part most migration posts get wrong, so read the table before you promise your team a one-click move.

What you have in Claude	Transfers automatically?	How
Claude Code project instructions and settings	Yes	Desktop app import, or /import in Codex CLI
Claude Cowork setup, skills and plugins	Yes	Desktop app import
Cursor configuration	Yes	Desktop app import, or /import in Codex CLI
Recent work and projects from those surfaces	Yes	Selected on the "Select items to import" screen
claude.ai chat Projects and their knowledge files	No	Manual export and rebuild
claude.ai custom instructions	No	Copy and paste into ChatGPT custom instructions
Claude memory entries	No	Export, review, paste the ones worth keeping
Connectors and MCP servers	No	Reauthorise each one in ChatGPT
Conversation history from claude.ai	No	Data export, kept as an archive

The short version: your agentic and developer surfaces move natively. Your chat surface does not. If your marketing team's "agents" are really claude.ai Projects with a system prompt and eight PDFs in the knowledge base, budget an afternoon per Project to rebuild them. If your agents are Cowork skills, the import does the heavy lifting.

How to import Claude agents into ChatGPT: the desktop app route

This is the main path for business teams. It requires the ChatGPT desktop app, not the browser. OpenAI's own documentation lives at learn.chatgpt.com/docs/import, and it is worth having open on a second screen while you work.

Step 1: Install and sign in to the ChatGPT desktop app

The import surface only exists in the desktop client. Sign in with the work account that holds your Business or Enterprise seat, not a personal one. If you sign in personally, you will import your work setup into a consumer account with different data terms, and you will have to redo it.

Step 2: Open Settings, then Import

Go to Settings and look for Import. If Import is not showing as its own section yet, open General and find Import other agent setup. Rollouts have been staged, so the location varies by version. Update the app before you conclude the feature is missing.

Step 3: Select Import and choose your sources

Select Import. You will be shown the agents ChatGPT can detect on the machine. Currently that means Claude Code, Claude Cowork and Cursor. Tick the ones you want to pull from, then select Continue.

Note that detection is local. ChatGPT reads what is installed and configured on that computer. If your Cowork setup lives on a colleague's laptop, the import has to run on their machine, under their account.

Step 4: Choose what to bring over

On the Select items to import screen you pick from instructions, settings, skills, plugins, projects and recent work. Then select Continue.

Resist the temptation to tick everything. This is your one clean opportunity to leave behind the prompt somebody wrote in a rush in February. Import the nine things from your inventory. You can always run the import again later.

Step 5: Open an imported project and test it

When the import finishes, open an imported project or chat and carry on working. Do not treat this as done. Run your three highest-value tasks through the imported setup and compare the output against what Claude produced last week. Different model family, different behaviour, even with identical instructions.

Step 6: Decide on automatic updates

In Settings, Import you can turn on automatic updates so imported work stays in sync with the original agent. Turn this on if you are running a parallel period where people still work in Claude. Turn it off on the day you cut over, otherwise you will spend a month unsure which tool holds the current version of anything. Your import history sits in the same settings section if you need to check what came across and when.

How to import into Codex CLI

If your engineers live in the terminal, the route is shorter.

1. Start a local Codex CLI session in the repository you care about.
2. Type /import.
3. Choose Claude Code or Cursor.
4. Confirm the items you want and let it run.

Run it per repository rather than once globally. Project-level instructions are where the real institutional knowledge sits, and importing them repo by repo gives your team a natural moment to review what each file actually says. Plenty of them will contain a rule that stopped being true in March.

We covered the wider Codex picture, including where it beats and loses to Claude Code, in the OpenAI Codex 2026 update guide and in when to use Claude AI vs Copilots vs Claude Code.

The manual route: claude.ai Projects, Skills and memory

This is the work the import button does not do, and for most non-technical UK teams it is the majority of the migration.

Projects

For each claude.ai Project worth keeping:

1. Open the Project and copy its custom instructions into a text file.
2. Download every file in the Project knowledge base.
3. Create the equivalent ChatGPT Project.
4. Paste the instructions in, adapting them to ChatGPT's phrasing conventions rather than pasting verbatim.
5. Upload the knowledge files.
6. Run three real tasks through it and compare against the Claude output you saved.

Instructions written for Claude tend to be longer and more explicit about reasoning. ChatGPT generally responds better to tighter, more directive instructions. Paste first, then cut roughly a third. If you want the pattern we teach for writing these properly, it is in from prompting to task delegation, and there is a build guide in building custom GPTs for enterprise teams.

Skills

Cowork skills come across in the desktop import. claude.ai Skills do not. If your team invested heavily here, and some have, this is your biggest single rebuild. We wrote about the scale of that investment in I stopped prompting, I built 50 Claude Skills instead. The honest read: a mature skills library is a genuine reason not to switch, or at least to keep a small number of Claude seats for the people who depend on it.

Memory and custom instructions

Export your Claude data from Settings, Privacy. Open the memory entries, read them, and move only the ones that describe how you work rather than what you were doing in April. Most teams keep about a fifth. Paste those into ChatGPT's custom instructions and personalisation settings.

Connectors

Every connector reauthorises from scratch: Google Drive, Gmail, Microsoft 365, Notion, SharePoint, your CRM. Nothing carries over, and this is usually where the migration stalls because it needs IT rather than the team. Raise the tickets in week one, not week three.

The UK compliance work nobody schedules

Changing AI vendor is a change of data processor. Under UK GDPR that is not a procurement detail, it is a documented event. Four things need doing, and they are the reason a two-hour migration turns into a two-week one.

1. New Data Processing Agreement. Your Anthropic DPA does not cover OpenAI. Get the OpenAI DPA executed before staff put client data into the tool, not after.
2. Update your Record of Processing Activities. The ROPA names your processors. If it still says Anthropic in November, that is a finding waiting to happen.
3. Refresh or redo your DPIA. For a firm-wide rollout touching personal data, the ICO expects a Data Protection Impact Assessment. A DPIA written for Claude does not transfer to ChatGPT because the processing, retention and sub-processor picture is different.
4. Check data residency and retention. Where the data sits, how long it is kept, and whether it can be used for training are all plan-dependent on both sides. We went through the questions to ask in AI and data residency for UK enterprises.

There is one more thing worth flagging because it bites during migrations specifically. When people are moving between two tools, they share and copy far more than usual, and shared conversation links leak. We set out the guardrails in the risks of shared AI conversations. Send that to your team before the migration, not after.

"The migration risk is never the file transfer. It is the fortnight where half the company is in one tool, half is in the other, and nobody is sure which version of the client deck is current. Set a hard cutover date on day one and announce it. Ambiguity costs more than either licence."

Toni Dos Santos, Co-Founder, Spicy Advisory

A five-day migration plan for a UK team

This is the plan we run with clients between 50 and 1,000 employees. Compress it for a team of ten, extend it for anything over 500.

Day 1: Inventory and decision

List every Claude asset, owner, surface and criticality. Pull three months of actual spend by product. Confirm the switch still makes financial sense once rebuild time is priced in at real day rates. Roughly a third of teams stop here, and that is a good outcome.

Day 2: Procurement and compliance

Buy ChatGPT Business or open the Enterprise conversation. Execute the DPA. Start the DPIA. Raise IT tickets for every connector you will need. Do not skip to day three because this is the slow part.

Day 3: Import

Run the desktop import on every machine that holds a Cowork or Claude Code setup. Run /import in Codex CLI for each active repository. Turn automatic updates on. Rebuild the top three claude.ai Projects.

Day 4: Test against real work

Each team runs its three highest-value tasks in both tools and scores the output. This is the only evidence that matters, and it is the thing that stops a migration turning into a six-month argument. Rewrite the instructions that underperform.

Day 5: Train and announce the cutover

Ninety minutes of hands-on training per function, on their workflows, not on features. Announce a hard cutover date two to three weeks out. Keep Claude seats live until that date, then cancel. We cover why workflow-specific training beats tool tours in the four-phase enterprise AI adoption framework, and there is a real example in going from 18% to 72% adoption in eight weeks.

Should you actually switch? The honest maths

A migration is not free just because the import button is. For a 60-person team, price it properly:

Line	Realistic cost
Annual licence difference (Claude Team to ChatGPT Business, 60 seats)	Often close to zero at list price
Rebuilding claude.ai Projects and Skills	3 to 10 days of someone senior
Compliance work (DPA, ROPA, DPIA)	2 to 5 days, legal or DPO time
Connector reauthorisation	1 to 3 days of IT
Training and the productivity dip	2 to 4 weeks of partial output

If the saving is a usage line that swings by thousands a month, the maths works comfortably. If the saving is a few pounds per seat, it does not, and you are better off fixing usage governance where you are. There is a full UK cost picture in how much AI training costs in the UK and in the Claude Opus 5 business guide if you want the other side of the argument fairly stated.

One more option that gets overlooked: you do not have to go all in. Keep six Claude seats for the people whose work genuinely depends on a mature skills library, move the other 54 to ChatGPT, and you capture most of the saving without destroying the capability you spent a year building. Nobody sells this option to you because it is not a clean win for any vendor, but it is frequently the right answer.

And if the direction of travel is the other way, we have the mirror-image guide: how to switch from ChatGPT to Claude or Gemini without losing your data. There is also Claude.ai 101 for business teams and the ChatGPT Work and GPT-5.6 guide if you want each platform on its own terms.