Engineering guides and deep dives on Optimization - short how-tos and long-form breakdowns.
POST /v1/decisions on gpt-6-luna costs $0.10 per million input tokens with no cache-read, cache-write, or output charge, and one 1,000-trial test saw a choice question return heads 98% of the time on a 70% coin.
Anthropic's minimum cacheable prefix runs backwards: newer models have the lower floor, so one prefix caches on Opus 5 and silently never on Haiku 4.5.
Fast mode buys up to 2.5x speed on Opus for a flat ~2x per-token premium with no quality change; lower effort cuts tokens instead. Two controls that stack.
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 plus 2–3 generic agent metrics - about five. BLEU/ROUGE miss semantics; G-Eval, DAG, and QAG cover what humans 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 solve-rate drop, per JetBrains.
Prompt caching cuts 90–97.5% off repeated tokens, batching cuts 50%, and the two stacked cut 95% or more. A 2026 PwC study measured 41–80% real savings from caching alone.