Pay for the quality you get

Most APIs quote one price for a model and never tell you which version of it answered. We quote a price per build, and the bill follows the weights that actually ran.

One model, a ladder of prices

Qwen3.8 27B, per million output tokens:

Build	Quality	Price
UD-Q5_K_M	reference	€0.75
UD-Q4_K_M	reference	€0.62
UD-IQ4_XS	balanced	€0.54
UD-Q3_K_XL	balanced	€0.50
UD-IQ2_XXS	compact	€0.28

The advertised price is what the best quality costs, so you are never billed above it. Everything else is a discount.

Set a floor and stop thinking about it

curl … -H "X-ElephantPool-Require: quality:reference"

One header and your requests only ever reach machines serving reference weights. It is a floor, not an exact match: better quality is always accepted. Set it and your price is predictable again.

Leave it off and you get the best build that is free at that moment, at that build's price — usually cheaper, occasionally the best one for less than a competitor charges for their worst.

Why the cheap end is genuinely cheap

Smaller weights decode faster, so a host serving them produces more tokens an hour. Pricing by build passes that through to you instead of pocketing it. At €0.28 per million tokens, a 27-billion-parameter model costs less than most of the market charges for something a third the size.

And why we can prove it

Every completed request comes back with a receipt signed on the machine that served it, naming the sha256 of the weights file it loaded. The discount is applied from that signature — not from anything the host tells us. A host cannot serve the cheap build and be paid for the expensive one.

See what a receipt proves and what you pay, and why it moves.