Kimi API Billing: Token Usage, Cache Writes, and Consumer Membership Are Different

The subscription boundary is explicit

Moonshot’s official billing overview says Kimi API inference is pay-as-you-go and is separate from the Kimi membership and coding products. A team purchasing the consumer app for researchers should therefore budget production API use on another line, with its own project keys and spending alerts. The company also explains that tax is determined by jurisdiction at checkout; listed US-dollar API reference rates are not an assertion of the user’s final local invoice total. Avoid reusing a billing assumption from a domestic CNY page without also confirming the account’s payment region.

Cache read and cache write are separate behaviors

The official pricing guide describes automatic cache handling for repeated prefixes, with a default five-minute time to live, and different write time-to-live choices on supported K3 models. A cache miss may incur a write cost, while a later cache hit is charged at its published hit meter rather than again as an uncached input. There is no reason to assume every repeated query hits: changes in shared system instructions, attachments or tools can invalidate the prefix. Measure cache events from returned usage fields; do not manufacture hit rates to make a forecast look cheaper.

A reproducible four-meter example

The provider’s international homepage lists the K3 model at USD 3 per million input tokens, USD 15 per million output tokens, USD 3 per million cache writes and USD 0.30 per million cache-hit tokens. For a hypothetical batch with 300,000 new input, 100,000 written cache, 700,000 cache hits and 200,000 output tokens, a simple line-item sum would be USD 4.41 if those are independently billable categories: 0.3×3 + 0.1×3 + 0.7×0.3 + 0.2×15. Do not add a write charge if the provider usage record already accounts for the same tokens under a mutually exclusive bucket; the event definitions control the bill.

Version status affects whether a rate is useful

Moonshot’s model catalog includes discontinuation notices for older Kimi series, and its billing guide provides a planned October 31, 2026 restriction for K2 models in new accounts. A rate that was valid for a legacy endpoint does not make it purchasable by a new customer. The commercial record should carry at least an availability state and both the vendor’s effective/retirement dates when known. Avoid claiming that a named model is available today just because a retired version still appears in API examples.

Checklist before approving a production budget

First select an immutable model ID and confirm which portal provides it. Next obtain a rate quote by input, output, cache-hit and cache-write units, including time-to-live and any long-context threshold. Then test the account’s rate limit and required prepaid balance on a real staging workflow. Finally track taxes and FX outside the base USD token rate. This creates a meaningful recurring-cost forecast and cleanly hands model-quality evaluation to AI 02 rather than presenting API pricing as a performance ranking.

Decision checklist and limitations

Before purchase or deployment, retain the literal model or plan identifier, payer/merchant, country or cloud region, billing currency, exact usage meter, source URL, checked date, effective date if explicitly supported, and any plan commitment or promotional eligibility. Recalculate from the companion price and quota snapshots rather than assuming that a remembered list tariff will remain unchanged. A credit allowance, maximum input length, published token rate and confirmed billed currency answer different questions; never add them as if they were the same quantity.

When comparing another provider or route, use the same input/output distribution and acceptable result definition. Include unsuccessful attempts, optional tools, storage, tax, payment and human review only when those components are actually measured or supported by an explicit source. Vendor marketing and documentation are not independent product tests. This article makes no claim to have personally tested paid endpoints, purchased plans, or checked account-specific taxes or availability.