Deep diveProvider guide

Prompt caching

Mark the end of the stable prefix (tools, then system, then messages) with cache_control. Cache reads cost 0.1 times base input, lower on some models, and five-minute writes 1.25 times; check cache_read_input_tokens and cache_creation_input_tokens in usage.

From
Provider guide: published by Anthropic on its own site, platform.claude.com.
Checked
, against the source

Dossiers