Prompt caching
A feature that lets you mark parts of a prompt as 'cacheable' — subsequent requests reusing that prefix are served at ~10% of base input cost.
Prompt caching addresses a specific waste: every turn in a multi-turn chat re-sends the entire prior context (system prompt, prior messages, attached documents) to the model, and the model re-processes all of it from scratch. For long chats or heavy system prompts, that's a lot of wasted compute paying for the same tokens over and over.
The solution: mark the stable prefix as cacheable. The first request writes the processed representation to a short-lived cache; subsequent requests hitting the same prefix read the cache at a fraction of the base input cost. Anthropic prices cache writes at 1.25× base input (5-minute TTL) or 2× (1 hour), and cache reads at 0.1× — a 90% discount on the read.
For LADLE's typical multi-turn chat, that's roughly 40-60% savings on total input-token cost across a 5-turn conversation. The first turn pays the write premium; the next four turns benefit from the read discount.
Two things to know: (1) there's a minimum cacheable prefix size (1024 tokens for Sonnet, 4096 for Haiku), so short prompts silently skip caching; and (2) the cache is content-keyed, so cross-conversation reuse of a common system prompt is possible when many chats share the same shape.