Engineering guides and deep dives on Optimization - short how-tos and long-form breakdowns.
Anthropic's minimum cacheable prefix is model-dependent and runs backwards - newer, pricier models have the lower floor, so the same prefix can cache on Opus 5 and silently never cache on Haiku 4.5.
Fast mode buys up to 2.5x speed on Opus for a flat $10/$50-per-million-token premium with zero quality change, while a lower effort level cuts tokens - and sometimes quality - instead; they're independent controls meant to stack, not substitute for each other.
Prompt caching only pays when the prefix stays byte-identical - put stable content first, dynamic content last, and verify with cache_read_input_tokens.
Pick 1–2 custom LLM-judge metrics for the use case plus 2–3 generic agent metrics - about five total. BLEU/ROUGE miss semantics; G-Eval, DAG, and QAG cover what humans actually grade.
Output tokens cost ~5× input and never cache - cut free-form prose first. Rewriting inside a cached prefix mid-session can raise the bill, not lower it.
Managing what a coding agent keeps in context - not window size - cut agent cost over 50% on SWE-bench Verified with no drop in solve rate, per JetBrains Research.
Prompt caching cuts 90% off repeated tokens and batching cuts 50% flat - 95% stacked. A 2026 PwC study measured 41–80% real savings from caching alone on agent workloads.