The default mistake
When someone first uses AI seriously, they pick the strongest model. That is understandable: if the quality of the answer matters, why choose a weaker one?
The trouble is that this decision is made once and then never revisited. Six months later a thousand classifications a day are running on the top model that the cheapest tier would handle just as well — for a twenty-fifth of the price, and three times faster.
What distinguishes the tiers?
Not what most people assume. The difference between tiers is not knowledge — they all learned from the same vast pile of text. They differ in four things:
- Consistency on hard cases. All of them solve the easy task. The difference shows up on the edge cases: ambiguous, incomplete, unusual.
- Multi-step work. Where the task is five steps building on each other, the cheap tier loses the thread around the third.
- Handling long material. In a big context window the pricier tier uses information from the middle more reliably.
- Reasoning. Reasoning mode is typically available only on the upper tiers, and that is the one real capability difference.
How big is the difference in numbers?
Take a real example. 300 customer complaints a day to classify, roughly 800 input and 150 output tokens per complaint.
At 9,000 runs a month that is about 7.2 million input and 1.35 million output tokens. On the cheapest tier this is a few dollars a month. On the top tier the same thing is tens of dollars, and if it runs in reasoning mode, hundreds — because the thinking counts as output tokens too.
The amount does not sound big on its own. But this is one task. A company has twenty of them, and twenty times over is a budget line.
Cascading: cheap first, expensive only when needed
This is the best value-for-money pattern, and surprisingly few people use it. The idea: you are not choosing one model, you are choosing an order.
- First step: the cheap tier gets the task, and you also tell it that when it is unsure, it should write
UNSURE. - Second step: whatever came back unsure — and only that — goes on to the more expensive tier.
- Third step: whatever is still unsure there goes to a human.
In practice the large majority of cases are settled on the first step. The expensive model runs only on the fraction where it is genuinely needed — and the quality of the end result is the expensive tier's, not the cheap one's.
When speed is what decides
There are places where price is not the deciding factor but waiting is. If the user is sitting there watching, the response time is part of the product.
The cheap tiers are not only cheaper, they are substantially faster, and reasoning mode can be an order of magnitude slower. For an automatic suggestion shown in front of a customer, that difference runs between usable and unusable.
The reverse is also true: if the task runs overnight in a batch, speed does not matter — and at many providers a batch run costs half.
How to decide: the forty cases
You do not need an engineer for this, you need an afternoon.
- Collect 40 real cases from your own work. Not the prettiest ones: the real distribution, hard cases included.
- Write down what the good answer is for each one. This is the most important step, and it is the one people skip.
- Run all 40 on the cheapest and the middle tier. If the difference is material, run the top tier too.
- Count the errors, not the impression. "Seems better" is not data.
If the cheapest tier gets 38 out of 40 right and the top model gets 39, the question is no longer which is better but whether one case is worth paying twenty-five times the price for — and the answer depends on the task, not on the model.
Where the cheap tier is genuinely weak
So there is no misunderstanding: there are places where you should not economise.
- Where an error is expensive. Contracts, legal text, medical content, financial calculation.
- Where many rules have to hold at once. The cheap tier drops rules sooner — and does not mention it, as Constraints That Stick explains.
- Where several steps build on each other. Across an agent's ten steps, small errors multiply.
- In smaller languages. Performance in Hungarian and Spanish varies far more on the cheap tiers — that has to be measured, not assumed.
What to take away
Picking the tier saves more than switching vendor — inside one family the price gap is twenty-five fold, and the cheap tier handles most everyday tasks. The good pattern is not to pick one but to cascade: cheap first, expensive only for the uncertain cases. And the decision is settled on forty real cases, not on a feeling.