DEV Community

TapixAi
TapixAi

Posted on

Why we publish the models where we are more expensive

Every price comparison we publish includes the models where we are more expensive than the official rate.

That sounds like a strange thing for a reseller to do, so here is the reasoning — and the numbers behind it.

The pitch we did not make

We resell API access: one OpenAI-compatible endpoint, many models, prepaid credit. The obvious marketing line is "everything cheaper than list". It is also false, and anyone can check it in five minutes:

  • On some models our price sits below the official rate, because of how we buy capacity.
  • On others it sits above it, usually because our upstream charges more for that route than the vendor's own direct price.
  • On a few it is roughly equal.

The distribution is not stable either. Upstream pricing moves, so a model that is 12% below list today can be 4% above it next month.

Why we publish both columns anyway

Three reasons, in order of how much they matter to us:

1. The only useful comparison is per model.
A single "up to 70% cheaper" claim tells you nothing about your actual bill. If you need one specific model and we are 3% above list on it, the honest answer is "use the vendor directly for that one".

2. A table you cannot audit is marketing, not data.
Every row on our pricing page can be checked against two public numbers: our price, and the vendor's list price. When a reader finds a row that does not add up, that is a bug, and we would rather hear it than hide it.

3. It changes what we build next.
The models where we are above list are exactly the ones where our supply agreement is bad. Publishing that number keeps it in front of us instead of burying it in a spreadsheet nobody reads.

How the numbers are produced

Roughly every six hours a job measures our real cost per model: it sends a fixed probe, reads the upstream billing counter before and after, and divides by the tokens actually reported. If the sample is too small or the upstream returns garbage, the measurement is marked low-confidence and automatic repricing is frozen for that model instead of guessing.

Then a pricing floor applies: if the measured cost would push net margin below 20% after fees, the price goes up. Prices never drop automatically — a discount is a decision, not an experiment.

What this means if you are comparing providers

Two practical consequences:

  • Do not trust a comparison table without a "we are more expensive here" column. If a reseller only ever shows favourable rows, assume the rest exist and are hidden.
  • Compare on your own traffic. Ten real prompts, one week, two invoices. That beats every table on the internet, including this one.

Our full price list, both columns, is on the pricing page. If a row looks wrong, email support — we would rather fix it than have it sit there flattering us.


We run Tapix AI, an OpenAI-compatible API. $1 of free credit on signup, no card required to test.

Top comments (0)