What each one costs
Cost is not one number. Money is only the part that turns up on a bill, which makes it the part people notice last and the part that decides least. Set the two side by side across everything they actually cost you and the picture is not that one is cheaper than the other. It is that they lose on completely different lines.
| What it costs | A rule in code | The same rule asked of AI | Favours |
|---|---|---|---|
| Money | Effectively nothing, however often it runs. Written once, then paid for. | Tokens on every call, so the bill grows with every user and every run. | Code |
| Time | Answers in a fraction of a millisecond. | A wait on every call, measured in seconds. | Code |
| Accuracy | Exact, on the case you actually wrote. | Usually right, occasionally not, with nothing to tell you which. | Code |
| Policy | Provable. You can show the action cannot happen. | Best effort. You cannot prove it will never happen. | Code |
| Effort to build | You have to state the rule exactly before you can write it. That is the real price. | A sentence. That is the whole appeal. | AI |
| Effort to change | A code change, tested and deployed. | Reword it. | AI |
| The case you never foresaw | Unhandled. It does what you wrote and nothing more. | Handled, sensibly. That is the entire reason to reach for it. | AI |
The money line has a sting in it. An instruction in the prompt looks like a one-off, a sentence you add once and are done with. It is not. It is read again and charged again on every run that follows, so a line you added in a minute becomes a small standing charge on everything the system does from that day on. Multiply it across every run, every user and every month. The sentence that felt free is now a line on the invoice. That is usually where people find out.
Read the table downwards and its shape is the answer. Code wins everything you pay for on every single run: the money, the wait, the certainty, the ability to prove it. Claude wins everything you could not write down in advance: the case you never foresaw, the rule you cannot state, the change you make by rewording a sentence rather than editing code. So the question was never which of them is better. It is which of those two lists your problem actually lives on.
Part of When to use AI and when not to.