The four things that eat into Claude Code limits, in order

Claude Code pricing is shown in tokens, but the real cost is GPU inference time, and the per-token price is set by model size, direction (input or output) and whether the cache is used. Input (prefill) is processed in parallel while output (decode) is generated sequentially, so output costs about 5 times more per token; invisible thinking tokens are also billed as output, and effort and model settings carry over into the next session, so they need checking every session. Prompt caching costs up to 2 times the list price when writing and 0.1 times when reusing, which makes a difference of 20 versus 3.9 over a 20-turn conversation, but the cache breaks when you change the model or effort, compact, or let 1 hour pass. Nothing is sent only once — a file read once is re-sent every turn — so attach files with @-mentions, use /context to clear out unused MCPs, rewind if only the last few turns are the problem, compact if the whole session is heavy, and hand tasks with lots to read but short answers to subagents. In order of cost impact come long sessions, too much context, too heavy a model or effort setting, and broken caches; multipliers and policies change often, so re-check them against the official documentation.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.