Claude Code: Models, Cost, and Context

Every Claude Code session runs on top of three interlocking decisions you rarely make explicitly but that determine almost everything about how the session feels and what it costs: which model answers your requests, how much it's allowed to think before responding, and how much of the conversation it's carrying around in its context window at any given moment. Get these right and a session feels fast, cheap, and sharp. Get them wrong and you'll pay Opus prices for Haiku-shaped work, sit through minutes of "thinking" on a one-line question, or watch a four-hour session slow to a crawl as its context balloons.

None of this is mysterious once you see the mechanics. Claude Code exposes real, documented controls — /model, /effort, /fast, /advisor, /compact, /usage — and each one changes a specific, traceable thing about how a request is built and billed. This subject walks through all of them: the model lineup and how aliases resolve, how organizations restrict and substitute models, what effort levels and fast mode actually buy you (they are not the same lever), what the advisor tool does when Claude consults a second model mid-task, how the context window fills up and what survives compaction, how prompt caching keeps most turns nearly free — and exactly which everyday actions quietly throw that cache away. By the end, you'll choose a model/effort/context strategy on purpose rather than by habit.