Comparing costs
How token pricing actually behaves in production, and why the headline rate is rarely the whole bill.
Last updated 2026-07-24
Most models are priced per million tokens, split into an input rate and an output rate. Output is typically several times more expensive than input, which means the shape of your workload (not just its volume) determines the bill.
Estimate with realistic ratios
A summarisation workload reads a great deal and writes little; a generation workload does the reverse. Estimating both directions separately, using a real sample of your own traffic, is the difference between a forecast and a guess.
| Workload | Typical shape | Cost driven by |
|---|---|---|
| Summarisation | Long in, short out | Input rate |
| Chat assistant | Balanced, multi-turn | Both, plus history growth |
| Code generation | Short in, long out | Output rate |
| Classification | Short in, tiny out | Request volume |
Watch conversation history
In a multi-turn chat, every earlier message is re-sent as input on each new turn. Costs grow with the square of conversation length unless you truncate, summarise or cache. This is the single most common source of a surprising bill.
Models that aren't priced per token
Image, video, speech and search models use their own units: per image, per video second, per million characters, per thousand queries. The catalogue labels the unit next to every price, and comparisons never mix units silently: a row that cannot be compared shows an em dash rather than a misleading number.
Note: All prices shown across this build are illustrative placeholders for design purposes. Real rates will come from the provider catalogue at launch.