Buy seats when the work is varied and a person stays in the loop. Build a custom system when one defined task repeats at volume, has to reach into your own systems, and produces an output somebody can check against a rule. Price rarely settles the argument. Repeatability does. What follows is the arithmetic, at current list prices, and the decision criteria that sit behind it.
What is the actual difference between the two?
You are buying two different things. A subscription buys seats: a person opens an interface, types, reads the answer and decides what to do with it. A custom system buys usage: an event happens, a system you own sends work to the model, and something arrives in a place where the next step already lives. One scales with headcount. The other scales with volume of work.
That distinction matters more than any feature comparison, because it decides where the effort goes. With seats, the effort is training and habit. With a build, the effort is specification: writing down exactly what comes in, exactly what goes out, and exactly what counts as wrong.
What does each option cost?
A subscription is a flat, predictable per-user fee. A custom build has a one-off cost to specify and construct, then a running cost measured in tokens, which is usually far smaller than people expect. The table below uses Anthropic's published list prices for Claude plans and its API pricing, so you can put real numbers against both sides before anyone quotes you anything.
| Dimension | Off-the-shelf subscription | Custom system on the API |
|---|---|---|
| What you are buying | Seats, plus the apps included with them (Claude Code, Claude Cowork). | Usage, billed per million tokens, inside a system you own. |
| Headline price | Claude Pro $17 per user per month billed annually, $20 monthly. Claude Team standard seat $20 per user per month billed annually, $25 monthly, for organisations of 2 to 150 people. | Claude Opus 5 $5 in, $25 out per million tokens. Claude Sonnet 5 $2 and $10 on introductory pricing to 31 August 2026, then $3 and $15. Claude Haiku 4.5 $1 and $5. |
| How the cost grows | With headcount. Ten more people is ten more seats. | With volume. The Batch API takes 50% off input and output; cache hits cost 0.1 times base input. If you use Anthropic's hosted agent sessions rather than running the system yourself, those are billed at $0.08 per session-hour on top of tokens. |
| What starts the work | A person, remembering to do it. | An event: a new email, a new row, a schedule. |
| Integration | Whatever the app supports, plus copy and paste. | Whatever you build. |
| Time to first value | The same afternoon. | Weeks, and only after the task has been specified. |
| Where it stops working | The same task, done by hand, two hundred times a month. | Work that changes shape every time, or that nobody can define on paper. |
When should you choose the subscription?
Choose seats when the work is varied, judgement-heavy and person-triggered. If your team's AI use is drafting, summarising, researching, checking a document, thinking something through, then a subscription gives you all of that today for the price of a couple of hours of anyone's time, and nothing needs specifying first.
- The people who need it number in the tens, not the hundreds.
- The tasks differ every time, and a human judges the output anyway.
- Nothing needs to be written back into a system of record automatically.
- You cannot yet name the repeating task in a single sentence.
- You want evidence of value before commissioning anything to be built.
That last point is the strongest argument for starting here. A month of seat usage tells you which tasks people actually reach for the model to do, which is the input a build needs and the thing nobody can guess in advance.
When should you choose a custom build?
Choose a build when one task repeats with the same shape, arrives somewhere predictable, and produces an output you can check against a rule. The test is not whether the task is difficult. It is whether it is describable. If you can write down the input, the output and the definition of wrong, it can be built. If you cannot, no amount of budget will fix that.
- The task runs at least weekly and looks broadly the same each time.
- The input lands somewhere predictable: an inbox, a form, a folder, a CRM record.
- The output is checkable, so failures are visible instead of silent.
- The trigger is an event, not a person remembering on a busy morning.
- The work needs to happen when nobody is at a desk.
The economics follow from volume. Anthropic's pricing documentation works a public example of 10,000 support conversations at roughly 3,700 tokens each, run on Claude Haiku 4.5, coming to about $37 in total. That is the running cost of a system, not the cost of building or owning one, but it does show where the money in these projects actually goes: almost never on tokens.
Is there a middle path?
Yes, and for most UK SMEs it is the right one: subscribe first, watch what people use it for, then build the single thing that has already proved itself. This sequencing matters because the hard part is not the technology. The Office for National Statistics found large language models were the most widely used AI technology among UK businesses with 10 or more employees in June 2026, at 18%, and lists difficulty identifying business use cases, cost and a lack of expertise as the main barriers to adoption.
Read that carefully. The named barriers are about knowing what to do, not about access. Buying a subscription removes the access problem in an afternoon, and leaves the real one untouched. What closes the gap is picking one workflow, measuring how often it runs and how long it takes, and only then deciding whether it deserves to be a system. That is the whole logic of an AI Audit.
What do the numbers not tell you?
The price comparison hides more than it shows. The biggest omission is that running cost is not total cost: somebody has to review outputs, notice when the system is wrong and own it when it fails, and that time is real whichever option you pick. The rest is waste, and it lands on both sides of the choice. A subscription bought and never used costs the same as one used daily, while a build that automates a task nobody needed is a permanent liability rather than an asset.
There is also a genuine uncertainty here that no article can resolve for you. Nobody can tell from outside whether your repeating task is worth automating, because the answer depends on numbers you probably have not measured. What would settle it: count how many times that task ran last month, and how long each run took. If the total is under an hour, leave it alone. If it is days, you have found something, and a custom system starts to make sense.
Decide on repeatability and evidence, not on which option sounds more advanced. The subscription is not the beginner's version and the build is not the grown-up one. They solve different problems, and plenty of businesses should be running both.