Classification: provider-specific usage/pricing guidance. Check official documentation for the selected model and date; there is no universal cache definition.
View the full infographic → · View this article
When a provider reports cached input separately, preserve that field instead of adding it to ordinary input a second time. Cached input is still part of the request context; billing treatment and eligibility depend on provider and product.
Keep the counters distinct
- Uncached input: provider-reported input not classified as cached for that request.
- Cached input: provider-reported cache-read usage under the provider’s definition.
- Output: generated tokens, often priced separately.
- Cache writes/storage: may have separate pricing or timing rules and may be excluded from a simple estimate.
Do not infer a cache hit from prompt similarity. Record the fields returned for each attempt and compare equal workloads. A zero in one run means that run reported zero; it does not establish a model’s general cache behaviour.
Price sources and exclusions
The Context Cost Calculator links to official OpenAI API pricing, Anthropic pricing and Gemini API pricing pages. Presets are dated and model-specific. Its simple estimate excludes some cache-write, storage, tool/grounding, tax and discount charges; check the official schedule for a complete invoice forecast.
In the one bounded canary described in the request-boundary article, provider-reported cached input was zero for that run. It does not prove caching is unavailable or generally absent.