Prompt caching, measured

What to take away

• Caching is a prefix match: tools, then system, then messages. One changed byte invalidates everything after it.
• Put what never changes first and what changes every turn (time, user input, IDs) last.
• Check usage.cache_read_input_tokens in production. If it stays at zero, something in the prefix is changing.
• Writing to the cache costs a little more than a normal request, so caching pays off from the second hit.
• Caches are per model and expire after about five minutes of no use. Switching model starts cold.