Subscription or API key

For most models there are two ways to get access.

The first is an API key. You set up a tool, or your own code, to use a model. You paste in the key for that provider. From then on you’re charged on usage, per token. You typically buy credits in advance.

The second is a subscription. You pay a fixed amount every month for a product: a chat app, a coding tool, an editor. ChatGPT Plus, Claude Pro, Cursor Pro, that kind of thing. The model is included, within limits.

Paying for one usually doesn’t include the other. A chat subscription rarely gives you API credits. Check the provider’s current terms instead of assuming.

How subscriptions work in practice

Subscriptions have limits. How much you get depends on the plan, and providers don’t always tell you exactly. With Claude Pro, for example, there’s a usage window that resets every five hours. Hit the limit and you wait, or you upgrade to a more expensive plan.

The upside is predictable spending. If the AI stops because I used it too much, I switch to another tool, or I go walk the dog. It’s a fixed monthly cost and there are no surprises.

If I were heads down on a big project I really wanted to ship, I’d pay more to remove those roadblocks. The extra productivity pays for itself. But most days I’m not, and a mid-tier plan is plenty.

How API keys work in practice

With an API key, usage is metered and there’s no window. You can run as much as you can pay for. That’s what you want when the model becomes part of your software: classifying every support request, extracting data from uploaded documents, generating a draft inside your own product.

A few warnings from experience:

• Prepaid credits can expire, often after a year. Check the terms.
• Keep the key on the server. Never in browser code.
• Set a spending limit and a usage alert before you ship.
• Handle timeouts, rate limits, and provider errors.
• Validate structured output before you use it.
• Decide what happens when the model is down.

API cost is more than the price of one token. Add retries, long context, generated output, tool calls, and your own engineering time. A cheap model that needs constant correction can cost more than a capable one that succeeds once.

For a rough estimate, measure average input and output tokens on realistic examples. Multiply by expected requests. Add room for retries and growth. Then put a hard cap in place.

You can start for free

Cost doesn’t have to be a barrier. Most providers have a free tier or a trial, and the web chat interfaces are free to use for questions. Some CLI tools give you generous daily limits with a personal account.

Keep one thing in mind: nobody gives anything away for free. During free promotional periods, providers usually collect your prompts and code to improve the next version of the model. Read the privacy page before you paste in anything you’d rather keep private.

My advice: start with a subscription while you’re learning what AI can do for you interactively. Move to an API key when you have a repeated job, a clear way to measure success, and a reason to put the model inside software.