Cached AI Input: Preserve the Cost Category Before Choosing the Price

Cached ai token pricing: what to decide

An application reuses a long prompt and benefits from caching. If cost records collapse all input into one category, finance cannot explain the margin improvement or assess whether a pricing promise has been honored.

Provider cache behavior and eligibility vary. Consult current documentation and retain the reported categories rather than inferring cached quantities from prompt similarity. Commercial rates can differ from provider costs, but the offer should remain clear.

How this works in practice

In a hypothetical model, a million ordinary input tokens cost $2 and cached input costs $0.50. If 800,000 tokens are cached and 200,000 are ordinary, the provider input cost is $0.80 before any other categories. These are illustrative, not current vendor rates.

What can go wrong

A cache miss can change cost without changing the customer-visible action. Do not promise a fixed margin from caching alone. Validate performance and economics under both hit and miss scenarios.

Questions for your team

• Capture provider-reported cache categories.
• Separate cost rates from customer prices.
• Model cache misses as well as hits.
• Explain any customer-facing cache discount accurately.