Anthropic cache_read_input_tokens is 0 on every request: your breakpoint is on content that changes each time

Not specified by the original author.

Cache writes on every request, reads on none: move the breakpoint

Symptom

cache_creation_input_tokens is nonzero every request, cache_read_input_tokens is 0 every request. You pay the 25 percent write surcharge every time and never get a read.

Confirm the cause

Look at the block carrying cache_control. If anything in the prefix up to and including that block differs between requests, a timestamp, a request id, the new user message, the prefix hash differs and the lookback finds nothing. The docs call this out explicitly: cache writes happen only at your breakpoint, and the lookback only finds entries prior requests wrote, not stable content sitting behind the breakpoint.

Fix

1. Move cache_control to the last block whose prefix is identical across the requests you want to share the cache, usually the end of the static instructions or context, not the varying block.
2. If you used automatic caching (top-level cache_control), it lands on the last cacheable block, which is often the varying one; switch to an explicit breakpoint on the static prefix instead.
3. Keep the breakpoint byte-identical: any change at or before it produces a different hash.

Verify

The next identical-prefix request returns nonzero cache_read_input_tokens and near-zero cache_creation_input_tokens. Two consecutive requests with the same static prefix are the test.

The breakpoint-on-changing-content trap, the lookback that only finds prior writes, and the automatic-caching variant of the same mistake are documented on the prompt caching page. Verified live at https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching.