Hey folks,

There are now more than a dozen products that will sit alongside you and do actual work. Claude Cowork, ChatGPT Work, Kimi Work, Microsoft Copilot Cowork, Qwen Studio, Manus, Genspark, Perplexity's Computer. Alibaba's QwenWork went into international beta 4 days ago. Most of them landed this year.

So the question is moving from which one is best to which one do I commit to.

Every comparison I have read answers that with a feature table. I think that is the wrong test, and I want to give you a better one. It’s about how well you can leave.

Let's dive in..

Before the useful part, the boring part, because it is genuinely hard to find written down anywhere.

ChatGPT Work is free. OpenAI's launch page says Chat, Work and Codex "are available on every plan, including Free" in the desktop app, globally, Mac and Windows.

Claude Cowork has no free tier and no trial at all. The cheapest way to see it is Claude Pro at $20 a month.

Perplexity's Computer needs their $200 a month plan. The $20 tier does not include it.

Manus gives you 300 credits a day free. Genspark gives you 100, which is roughly one slide deck.

Qwen Work gives you 500 to start.

I could go on, and that is the problem. Many others will not publish their prices at all. Others will not say what a subscription actually includes before overage charges start. Some are unavailable in large parts of the world. And every number in it will be wrong next month.

Here is the thing about the entry price. It is cheap, it is knowable, and it is reversible. You can be wrong about $20 a month and fix it in an afternoon.

What is not reversible is everything the AI tool learns about you afterwards.

6 months in, your assistant knows how your business works, who your clients are, what you have already tried, how you like things written, which arguments you have already had with it. Nobody sells that to you. It accumulates. And the day you want to try something else, that is the thing you are actually deciding whether to abandon.

So the test I would run before committing to any of these is not "what can it do." It is "what happens if I want to leave."

There is no export tool for most AI work platforms, it’s mostly manual.

Here is what Anthropic officially documents. You prompt it.

That is the export tool. You prompt and copy the output to a file.

Google documents the same approach for moving off ChatGPT. Their import flow hands you a prompt to paste into the assistant you are leaving, and you carry the answer back.

2 of the largest technology companies on earth have looked at this problem, and the answer they shipped is copy and paste.

I want to be fair though. Anthropic did build the import side, and it works: Settings, Memory, Start import, paste. It is on every plan including Free. They also note it prioritises work-related content, so personal detail may not survive the trip.

Once you start looking, the pattern is hard to unsee.

Anthropic built an importer. Google built an importer, and Google's names ChatGPT and Claude specifically. Both will happily take your context off a competitor.

OpenAI has neither. No memory export, and no way to bring context in from anywhere else. You can export your chat history as a zip, which takes up to 7 days to arrive and expires 24 hours after it does. Your memory is not in it.

Microsoft is the most complex. In Microsoft 365, Copilot's memory of you is stored in a hidden folder inside your own Exchange mailbox. Your IT admin can pull it out with eDiscovery. You can export your custom instructions and that is it. The person who can get your memory is IT Admin.

And there is no standard coming to fix this.

Claude currently has 2 memory systems running side by side. If you are on the new one, your memories are at Settings > Memory, listed under Topics, and you can read and edit each one.

If you are still on the old one, they are somewhere else entirely, under Settings > Capabilities. And Anthropic's help page says that "until September 9, 2026, you will see the option to export your legacy memory."

After that, the export option goes.

I do not know which of those 2 you are on. Go and look, and if you find the second one, get it out this week.

I assumed there was a legal floor here.

In Europe, the right to data portability covers data you provided. The governing guidance specifically excludes data the company derived or inferred about you. Your transcripts are probably covered. The profile built from them probably is not. No regulator or court has ruled on it either way.

In California, an Attorney General opinion says internally generated data a business holds about you are personal information and must be disclosed on request.

In Australia, there is no right to data portability at all. There is a right of access, with no requirement that anything be machine-readable, and an explicit ground to refuse access to evaluative information a company generated internally.

So the strength of your claim on 6 months of accumulated context depends less on what you built and more on which country you were sitting in while you built it.

Everything the comparison tables measure is the cheap, reversible half of the decision. The expensive half never appears in them, because it does not exist yet on the day you choose. It accrues quietly afterwards when you want to try out another newer AI platform.

One thing that should make you feel better, though. I did the export.
It created me 6 files about my profile and 20 to 30% of what had piled up was still true. Most of it was out of date. The lock-in feels heavier than it probably is.

Which is an argument for running the test now while it is cheap, rather than in 2 years when you assume it is impossible.

5 minutes, free, on whatever assistant you already use most.

Most people have never seen this list. It is worth 5 minutes to find out whether yours is worth protecting or whether, like most, it is mostly out of date and you were never really trapped at all.