Kimi K3 cache costs: choosing a 5-minute or 1-hour TTL

A low cache-read price is not the whole bill

Repeated system instructions, documents and tool definitions can form a reusable prefix. Kimi K3 documents automatic caching with a default 5-minute TTL and a 1-hour option. Cache writes are billed separately, so a write followed by a hit cannot be priced using only the lower read rate. Source 1 Source 2

Caching changes reuse and billing, not the amount of information a task needs. Its value depends on the tokens actually hit, how often writes occur and the interval between requests.

What the TTL changes

TTL is the cache’s validity period. The pricing documentation says a hit within that period refreshes the TTL, charges the cached-input rate for the hit portion and does not add another write charge for that portion. Source 2

Five minutes may cover closely spaced requests; longer intervals can justify comparing the one-hour option. A longer TTL is not automatically cheaper. Include its write rate, observed hits and workload timing. Block-level processing also means identical text should not be assumed to produce a fully cached prefix.

Include writes in the estimate

Separate ordinary input, cache reads, cache writes and output according to the provider’s usage and billing definitions. With quantities expressed in millions of tokens, a budget can be organized as:

Cost = ordinary input × input rate + cache reads × read rate + cache writes × TTL-specific write rate + output × output rate

This is a billing breakdown, not a direct mapping to arbitrary SDK fields. Avoid counting the same input category twice, and include tools or other charges where applicable. A complete numerical price table could not be reliably verified for this review, so no guessed rates or fixed savings percentages are supplied.

Youjie’s generic calculator illustrates input, output and cache-read arithmetic. K3-specific write and TTL charges need separate accounting before treating the estimate as a complete bill.

Run a useful comparison

Use approved test material with a stable reusable prefix and put changing questions after it. Record the first request, repeats within the TTL and requests after expiry, retaining timing, usage and billing details. Compare one-hour settings using equivalent tasks and similar request patterns.

Do not freeze business information merely to improve hit rates. Update the content when facts change and then observe caching behavior.

Choose from actual request timing

Frequent conversations and occasional document jobs have different economics. Sample real intervals and compare complete costs for both TTLs.

Moonshot’s usage-based API billing is separate from Kimi product memberships and Kimi Code subscriptions. On Youjie, confirm whether the selected route exposes these settings and billing details; an intermediary need not use the same accounting as the official API.