Almost every model built in the last couple of years answers to the same shape of request. What you are choosing is a supplier, and you can change supplier next month for the price of a line of configuration. Everything below is a shopping trip, not a commitment, which is the single most useful thing to know before you spend anything.
There are three places to shop. The lab that made a model will sell it to you directly, by the word. A ChatGPT or Claude subscription is a different purchase: a flat monthly fee for an app, with no spend cap because there is nothing to cap. A host will rent you somebody else's model by the word, and because several hosts often serve the same model, that is where the price competition happens. And if a model has open weights, you can download it onto your own machine, where nothing you type leaves the room and nobody meters you.
One account is enough to begin. OpenRouter reaches more models with a single key than any other account we list, so you can try a lab's flagship and a downloadable workhorse in the same afternoon without signing up twice; the specialist hosts like DeepInfra are usually cheaper once you know which model you want. What any of it costs sits on each model's page, host by host, and the cheapest hosted models is the same figure sorted.
Three models follow, one for each door. Each links to its live page, where the price, the hosts and whether it fits your machine are all current — those move week to week, which is why they belong there and not in this sentence.
Where next: The catalogue · Run one on your own machine