One model, nine prices

Ask any inference API which version of a model just answered you. Almost none will tell you. The same name covers weights that differ by a factor of three in size and a great deal in quality, and you pay one price either way.

We publish nine prices for Qwen3.8 27B, from €0.75 down to €0.28 per million output tokens, and the bill follows the weights that actually ran.

You are never billed above the advertised price

The headline is what the best build costs. Everything below it is a discount, so the number on the page is a ceiling rather than an average you might exceed.

One header fixes the floor

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

Your requests then only reach machines running reference weights. It is a minimum, not an exact match — better quality is always accepted — so setting it makes your cost predictable without ever costing you an available host.

Why the cheap end is cheap

Smaller weights decode faster, so the machine serving them produces more tokens an hour. Pricing per build passes that through instead of keeping it. A 27-billion-parameter model at €0.28 costs less than much of the market charges for something a third the size.

And why you can check

Every completed request produces a receipt signed on the machine that served it, naming the digest of the weights file it loaded. The discount is computed from that signature. A host cannot serve the cheap build and be paid for the expensive one, and we cannot bill you for a quality nobody delivered.

The full pricing reference is in the knowledge base.