How the estimate works

  • Tokens = UTF-8 bytes ÷ bytes per token, rounded up. Real tokenizers differ by model; use the override for an exact count.
  • Input per call = other prompt tokens + context tokens, where context is the whole text, min(text, chunks × chunk size), or min(text, recall budget).
  • Cost per call = input × input price + output × output price, per million tokens.
  • Per month = cost per call × calls. Prompt caching, batch discounts and retries are not modelled.

The three scenarios, defined

Whole text
Everything is sent on every call. Simple, and limited by the context window.
Retrieved chunks
A retriever selects k passages. Cheaper; quality depends on retrieval.
Memory recall
A memory layer returns current facts with sources up to a budget. Cheapest per call; adds memory-service cost (see pricing).

Limits of this tool

This is arithmetic on your inputs, not a measurement of any provider or of Auld. Prices are example values. Quality differences between scenarios are not modelled.

Why bytes rather than characters?

Bytes are stable across scripts; tokenizers tend to track bytes more closely than characters for non-Latin text.

Where do I find my model’s price?

On your model provider’s pricing page; enter input and output rates separately.