What happens if the provider withdraws the model?

You move to whatever they offer next, and behaviour changes with it. This is the concrete form of the dependency, so it is worth asking a vendor how much notice they give, whether old versions run alongside new ones for a period, and whether your prompts and settings would carry across.

Can we audit what it does?

Not the model itself, which is the honest answer and sometimes the deciding one. What you can audit is behaviour: run your own set of cases, record the outputs and compare across versions. Where a regulator requires more than that, this arrangement may not clear the bar and it is better to establish that early.

Should we avoid it on principle?

There is no need to. It is a supplier relationship, and organisations handle those all the time by knowing what the exit looks like. Keep your prompts, evaluation cases and retrieved material portable, avoid building anything you could not move, and the dependency stays a commercial matter rather than a trap.

Glossary

Closed-source model

A closed-source model is one you can use only through its owner's service, with the trained file never published, so you reach it over a connection and never hold a copy of it.

In plain terms

You rent access rather than take delivery. The company that built it keeps the file, runs it on their machines, and lets you send questions to it. You get the capability and never the thing itself, which is unremarkable as a commercial arrangement and has consequences worth knowing before something important depends on it.

Why it matters

Most of the capability people are impressed by arrives this way, so treating it as the disreputable option would be an odd conclusion to reach. What it is is a supplier dependency, and the useful move is handling it the way your organisation already handles those. Know what happens if the price changes, if the product is withdrawn or if the terms move, and know whether you could go elsewhere without rebuilding everything around it.

How it works

The trained file stays with its owner and you reach it over a connection. That single fact produces every other property on this page: you cannot inspect it, cannot run it anywhere else, cannot keep it if the arrangement ends, and equally have nothing to install, nothing to size and nothing to maintain.

You get whatever the provider is running, including changes you did not ask for. A product can move to a successor model without changing its name, which improves things more often than not and occasionally alters behaviour a workflow was relying on. An evaluation set you can rerun is what turns that from a mystery into a fact you noticed.

Capability is generally the argument in its favour, and it is a real one. The largest and newest systems tend to appear here first, because whoever spent the money to train them has an obvious interest in being the route to them. Published alternatives have closed a good deal of that gap and have not closed all of it.

Being closed says nothing about what happens to your material. Retention, access and whether anything is trained on your inputs are governed by the terms of your plan, and business agreements are frequently very different from consumer ones inside the same product. That question is answered by reading rather than by the arrangement.

What renting the capability actually means

Both columns follow from one fact, which is that the file stays with its owner. That is why the arrangement offers the most capable systems with nothing to maintain, and why it offers no copy, no inspection and no guarantee that what answered you last month is what answers you today. Organisations deal with dependencies of exactly this shape constantly, and the ones that do it well are not the ones that avoided them. They are the ones that knew what leaving would cost before they needed to know. In practice that means keeping the things you built, your instructions, your evaluation cases and your own documents, in a form that is not welded to one provider, so that a price change or a withdrawal is a decision to make rather than a project to survive.

Seen in the wild

Use any of the widely known assistants, which is this arrangement and is where most people's impression of what AI can do was formed.
ChatGPT
Put an identical request to a closed model and a published one through the same routing interface, which is the only comparison that settles anything about the gap.
OpenRouter
Download a published family and run it yourself to find out what you would actually be giving up, on your own work rather than on a benchmark.
Qwen

Common misconceptions

People assume

Closed means our data is less protected.

In fact

The two questions are unrelated. What happens to your material is set by the terms of your plan, and a closed provider on a business agreement may well retain less than an arrangement you assembled yourself. Read the terms rather than inferring anything from the word.

People assume

It is always more capable.

In fact

Usually somewhat, and less than the marketing suggests, and not for every task. Published models have closed much of the gap for summarising, classifying, extracting and drafting, which is most business work. The way to know is running both on your own material, which takes an afternoon and settles it properly.

Telling them apart