Token Usage & Cost Basics
Claude Code is billed by tokens (roughly ¾ of a word each). What you send and receive both count.
A token is roughly three-quarters of a word, so 1,000 tokens is about 750 words — whether that's your prompt, a file Claude Code reads, or the response it writes back. Every one of those counts toward usage, and pricier, more capable models cost more per token.
Correction: the video points to /cost and /stats to check usage. In current Claude Code, that's consolidated into a single /usage command, which shows session token/cost totals, a plan-usage breakdown (attributed to skills, subagents, plugins and MCP servers), and prompt-cache statistics. There's also /insights, which analyzes your recent sessions and writes an HTML report on how you actually work — not just how many tokens you burned.
Real-life use cases
Spotting an expensive habit
Running /usage after a slow week shows that a single MCP server is responsible for a disproportionate share of tokens, prompting the user to disable it when it isn't in active use.
Choosing model by budget, not habit
Understanding that output tokens on a more capable model cost more leads a user to default routine tasks (renaming files, small edits) to a cheaper, faster model and reserve the expensive one for genuinely hard problems.
Source
- Source video: “Claude Code Explained” — 27 concepts in 27 minutes (YouTube)https://www.youtube.com/watch?v=ZlDnsf_DOzg
- Manage costs effectivelyhttps://code.claude.com/docs/en/costs