Saying three things at once should not cost you three

For a while, our pricing counted every item the AI pulled out of what you said. Say three things in one breath and it cost three. It seemed obviously fair: you got three things, you paid for three things.

We reversed it. One submission is one unit now, however much comes out of it. Three reasons, and the first is the one that actually settled it.

A price you cannot quote in advance is not a price

Here is the problem with per-item, stated plainly: nobody knows how many items are in a sentence until the model has read it. Not you, and not us.

So at the moment you press the button, the cost is unknown. That leaves a system with two options, and both are bad. It can refuse the request because it cannot price it — refusing you for something you have not done yet. Or it can run anyway and let the balance go negative, which means the limit was never a limit.

There is no third option. You cannot show someone a price before an operation whose price is determined by the operation.

It taxed the one thing we tell people to do

The line on the front of the site is "You say it. We'll plan the rest."

Per-item pricing means saying three things at once costs three times as much as saying one. The feature the product is named for, the thing the whole capture pipeline was built to handle — multiple intentions in a single breath — was the most expensive way to use it.

You would not have noticed the mechanism. You would just have found yourself, after a few weeks, saying one thing at a time. And that is the version of the product that is worse in every way, arrived at by an accounting decision nobody argued for.

The cost was never the constraint anyway

The third reason is the least interesting and the most clarifying.

We measured what an extraction actually costs to run. It is a fraction of a cent. A free account's entire monthly allowance costs us about half a penny.

So the allowance was never recovering a cost. It exists to mark the line between tiers — and a limit that exists to mark a line should be counted in whatever unit is clearest to the person it applies to. That unit is the number of times you talk to it, not a number produced afterwards by a model neither of us watched.

What did not change

The tiers did not move. The prices did not move. The allowance numbers did not move. What changed is the unit they are counted in, which means everyone got more for the same money without a single figure on the pricing page being edited.

Voice is unchanged too: a spoken capture uses voice minutes for transcription, plus one unit like anything else.

The general form

The reason to write this down is not the pricing. It is the test that caught it.

Can the price be stated before the thing happens? If not, the model is wrong, regardless of how fair it looks on a spreadsheet afterwards. Fairness computed after the fact is not something anyone can plan around, and a limit you cannot plan around is just a surprise waiting to happen.