Prompt caching

Understand when a provider can reuse repeated model input and how to read the resulting token and cost fields.

Recurring translation requests often share instructions even when their source text changes. Prompt caching lets a model provider reuse a repeated input prefix instead of processing that input from the beginning each time. The request still generates a new translation.

Lingo.dev places stable instructions before request-specific material and supplies cache hints and routing appropriate to the provider. The provider decides whether a request receives a cache hit.

What can remain the same between requests #

System instructions, matching rules, and selected voice contribute to a byte-stable prefix while configuration is unchanged. Retrieved glossary material and request content follow that stable portion. Editing rule or voice text changes the prefix and can prevent reuse of the earlier cached input.

Supported Anthropic and Google paths use explicit ephemeral cache breakpoints. Other providers may cache automatically or use different behavior. Minimum input size, lifetime, and pricing are provider-specific; support for one does not establish the same behavior for another.

Repeated traffic with the same prefix can benefit from caching. Sparse traffic, changed context, a different provider route, or provider policy can require fresh processing. Stable configuration alone does not guarantee a hit, discount, or latency improvement.

Read usage without counting the same input twice #

Account token reports can include total input, output, cache-read, and cache-write tokens. Cache counters describe components of the input work; do not add them to total input as though every counter represented separate tokens.

For example, 1,000 total input tokens with 800 cache-read tokens do not mean 1,800 input tokens. Provider definitions determine which fresh and cached quantities can be derived from the report. Missing usage is not evidence of no work.

For an individual translation, open its model requests. The view shows recorded input and output tokens and provider cost for the full attempt, which can cover several document targets. The job response supplies progress and download URLs; use account reports to investigate cache activity and spending across requests.

Separate savings from other work #

A cache hit can reduce provider input processing while output generation and a localization charge still occur. Independent measurements, professional review, and the platform plan have their own costs. Use Usage for account spending and cost evidence for individual work.

Prompt caching also differs from translation reuse, which retains target text aligned with unchanged source. Context retrieval selects instructions for new generation. Keeping these mechanisms separate makes it possible to explain both a lower-cost request and a run that did not need to translate unchanged content at all.