What you are actually paying for, on a $20 AI subscription
A flat monthly AI subscription charges you the same whatever you ask it. That is the whole proposition, and it is a reasonable one — until you look at what you actually ask it.
Go back through a week of your own messages. Most people find something like this:
- Fix the grammar in this paragraph
- What does this error mean
- Summarise these notes
- Write a formula that does X
- Rewrite this so it is shorter
- Actually think hard about this architecture decision
Only the last one needs the most capable model in the world. The rest are the kind of work a model costing a fiftieth as much answers just as correctly. On a single-model subscription they all cost the same, because there is only one model.
The gap is bigger than people expect
Frontier models are priced per million tokens, and the spread between the cheap ones and the expensive ones is not ten or twenty per cent. It is one to two orders of magnitude. A short rewrite that costs a fraction of a cent on an everyday model can cost twenty or thirty times that on a frontier one — for an answer that is, for that particular task, indistinguishable.
That is not an argument for only using cheap models. A cheap model asked to design a schema, debug a race condition, or reason carefully about a contract will produce something confident and wrong, which is more expensive than any subscription. It is an argument for using the right one each time.
Why nobody does this by hand
You could, in principle, keep several subscriptions and pick the right one per question. In practice nobody does, because it means paying for all of them, remembering which is which, and making a cost decision before every message. The friction is larger than the saving.
So the choice on offer is usually: pay frontier prices for everything, or use a cheaper assistant and accept worse answers on the hard questions.
The third option
Score the task first, then route it. How long is it, how much reasoning does it need, what kind of work is it — and send it to the model that does that best for the least. Keep the expensive models for the questions that are actually hard.
That is what Switchboard does, and the part that matters is that you can see it: every answer shows which model wrote it, which tier it was, and what it cost. Not a monthly bill with no breakdown — a per-answer receipt. If the routing makes a bad call you can see that it did, and say so.
What it comes to
Switchboard Lite is $4.99 a month and reaches the same frontier models as the $20 plans, because it only sends them the work that needs them. Plus is $9.99, Pro $19.99, Builder $49.99, and the difference between them is how much you can use and how many of your own tools you can connect — not which models you are allowed to talk to.
If that sounds like it should be worse, test it rather than take our word for it. There are a few free questions on the front page with no account and no card: ask it the hardest thing you asked an assistant this week, and see which model it picks.
Switchboard routes each task to the model that does it best for the least. Ask it something — a few questions free, no account — or see what it costs.