All capabilities Cost and throughput

Stop paying twice for the same file

Mark the stable part of a prompt as cacheable and every later request reads it at 0.1x the input price instead of paying full price again.

DocsOfficially documentedTestNot tested by benchr Quick · API Docs checked September 1, 2026

Claude APIClaude Opus 5Claude Sonnet 5Claude Fable 5+3

Overview

Charges 1.25x base input to write a five-minute cache entry, 2x base to write a one-hour entry, and 0.1x base to read either.
Applies to the part of the prompt that does not change - the system prompt, the tool definitions, the document, the examples.
Reports what actually happened in the usage block, so the saving is measurable rather than assumed.

01Move everything stable to the front of the prompt and put the variable part last. The cache is a prefix.
02Mark the boundary with cache_control, and add ttl: "1h" only when the same prefix is reused less often than every five minutes.
03Check the minimum. Claude Opus 5, Fable 5 and Mythos 5 need 512 tokens; Opus 4.8, Sonnet 5, Sonnet 4.6 and Sonnet 4.5 need 1,024; Haiku 4.5 needs 4,096.
04Read cache_creation_input_tokens and cache_read_input_tokens on the response. Two zeros mean the prompt was under the minimum - no error is raised.

Short prompts silently do not cache. The API returns no error, just zeros.
A one-hour entry costs twice base input to write, so it only pays off if the prefix is genuinely reused.
Any change inside the cached prefix invalidates it - reordering tool definitions is enough.
The saving is on input tokens only; output is unaffected.

benchr has not run this capability. No test has been published.

benchr test runs →

Source-check history

September 1, 2026DocsOfficially documentedFirst logged. Multipliers, the two TTLs and every per-model minimum read from the caching guide.