Start with the tokens a request actually uses
An API request can have input tokens, output tokens, and eligible cache tokens. Input is the instructions and context you send. Output is the response you receive. Cache applies only when the request is eligible for the response cache.
Do not estimate a whole project from one prompt. A support bot, document extraction job, and coding assistant can have very different input-output ratios.
Write down one realistic scenario before you calculate
Choose a live model, then record the expected number of requests, average input tokens, average output tokens, and any repeated context that may be cacheable. Keep the scenario small enough to compare with a real usage record after your first test.
Use this outline: requests × ((input tokens × input rate) + (output tokens × output rate) + (cache tokens × cache rate)) ÷ 1,000,000. The result is only an estimate until the request is metered.
Use the live catalog as the price source
The EidosStack catalog shows the current IDR price per one million input, output, and cache tokens for available models. It is the source to check before a request, rather than a fixed number copied into an article.
Model supply can change. If a model is unavailable, choose another model deliberately and repeat the scenario with that model’s current card. Set a per-key budget before sending production traffic.