Almost nobody asks what AI costs before they have started using it. The question arrives later, usually from someone who has quietly accumulated four subscriptions and wants to know which ones to keep.
That is the useful version of the question, so it is the one worth answering.
There are three different prices, and most comparisons mix them up
A flat subscription. Around £20 a month buys the consumer tier of almost every assistant. The number is so consistent across vendors that it has stopped being a pricing decision and become a convention.
Metered per-token billing. You pay for what you send and what comes back, measured in tokens rather than words. This is what developers pay and what anything built on an API costs.
Free, with a ceiling. Not a trial. A real allowance that resets, with limits on volume rather than capability.
Comparing a £20 subscription against a per-token rate is where this goes wrong, because they are not the same product sold at different prices. One is an all-you-can-use allowance with a hard cap on the busiest day of your month. The other has no cap and no floor.
The subscription stack is the actual cost
The searches people type are about one tool. The bills people complain about are about five.
Five tools at £20 is £100 a month, or £1,200 a year, and almost nobody decides to spend that. It accumulates one reasonable decision at a time: the assistant, then the writing tool, then the image generator, then the transcriber, then the one a colleague recommended.
Nobody chooses to spend £1,200 a year on AI. They choose five £20 subscriptions, four months apart.
The uncomfortable part is that the overlap is enormous. A general assistant like ChatGPT, Claude or Gemini does a passable version of what most single-purpose writing and summarising tools do. The specialist is better at its one job. Whether it is £20 better every month is a different question, and it is the one worth asking before the renewal rather than after.
Per-token pricing: output is the expensive half
If you are paying by the token, there is one fact that matters more than the rates themselves: every major model charges more for what it writes than for what it reads. Often three times more. Sometimes five.
This is why so many cost estimates are wrong in the same direction. Paste a long document into a calculator and it will price the document — the input — and tell you a number that looks reassuring. The reply is where the money goes, and a request that returns a long answer can cost several times what it cost to ask.
We built the token counter specifically because of this. It counts both halves, prices them separately, and shows the date the rates were taken, because a pricing page that does not say when it was written is worse than no pricing page. It runs in the browser and needs no account.
What you are paying for is the ceiling, not the intelligence
This is the part that genuinely surprises people: on most consumer tools, the free tier and the paid tier run broadly the same model.
What the subscription buys is headroom — more messages, longer documents, priority when the service is busy, and the newest model first. Those are real benefits for someone who hits the limits. They are worth nothing to someone who does not, which is most people most of the time.
The free AI tools actually worth using goes through which free tiers are genuinely usable rather than a sales funnel with a time limit.
The only genuinely fixed cost is your own machine
Running a model locally costs nothing per use, forever. Ollama and LM Studio make this far less technical than it was a year ago, and open models like Llama and Qwen are now good enough at small sizes to be useful on an ordinary laptop.
The cost moved rather than disappeared: it is your hardware, your electricity, and a noticeably slower, less capable answer than the hosted version. For private documents and high volume it is unbeatable. For a difficult question asked twice a week it is a hobby. Running AI on your own machine covers what is actually involved.
So is it worth paying for?
Pay when a limit has already interrupted you.
Not when you anticipate hitting one, not because the paid tier lists features you can imagine using, and not because a free tier felt like a compromise. An interruption is evidence. Everything before it is a forecast, and forecasts are how five subscriptions happen.
How to choose an AI tool is the companion to this, and the n3os directory lists what each tool's free tier actually includes, in a write-up for every one of them, never ordered by who paid.