The monthly subscription is how most people meet this, and for heavy use of one lab's models it is a good deal. Paying per token instead feels like the risky choice, because an uncapped bill is a genuinely unpleasant idea to sit with. What the key buys in return is choice: any model on the market, the same key working in your editor and your tools as well as a chat window, and a quiet month that costs you nothing at all.
The reason the fear survives contact with the prices is that the unit is unfamiliar. Rates are quoted per million tokens, and a million tokens is a great deal of text. A paragraph runs to about a hundred. So the useful thing to do once, carefully, is count a heavy month.
Say a reply comes back at a few paragraphs, call it 400 tokens, and your question is 60. A twenty-turn conversation therefore has around 8,000 tokens written back to you. The reading is the larger and much less obvious half, because a chat app re-sends the whole conversation on every turn: by the twentieth question the model is reading the previous nineteen exchanges again, and the conversation totals somewhere close to 90,000 tokens read. Three conversations like that a day, five days a week, four weeks, and a heavy month lands at roughly five million tokens read and half a million written.
Both rates sit on the cards below, live, quoted per million. Five of the first plus half of the second is your month. Put that beside what the subscription takes from you each month and the comparison is one you have made yourself, which is the only version of it worth trusting.
Cap it in either case. Every serious provider lets you set a monthly spend limit, and a router gives you one limit across every model behind it. Set the cap to what the subscription used to cost and the worst case is that you have paid what you were already paying.
The honest ledger of what you give up is mostly about the app rather than the model. A subscription bundles memory of past chats, file handling, voice, image generation and image understanding, and a mobile experience somebody has polished for years. Going metered means choosing a chat front-end that takes an API key, and the good ones are genuinely good, but it is a setup step with an afternoon in it. Some people land on both, keeping a cheap subscription for the comforts and using the key for everything else, and that is a real answer.
The switch itself is three moves. Make a key at a provider or a router, set the spend cap, then point a chat app or your coding tools at it. Most of the afternoon goes on choosing your first model, which is what the verdicts on each model page are there to shorten.
Where next: What is a token · What is a model router · Compare model prices