mcp tool ~118,986 tok ($36.80) [mcp tool]; Read ~142,110 tok ($35.85) [path]; Read ~134,547 tok ($25.53) [path] and 558 more
Bound the output at the source: limit/offset on Read, head/tail or a grep filter on Bash, or delegate the search to a subagent so only the conclusion lands in the main thread.
14 redundant file re-reads in session 4b13bf04
Est. saving ~$106.26 / Rs 9,351
[file] x8, [file] x3, [file] x2 (~33,847 tokens per read, re-read on every later turn once in context)
The file contents are already in context from the first read. Re-reading adds a second copy that you then pay for on every subsequent turn.
Session f92e6b60 spent 77% of its cost re-reading its own history
Est. saving ~$45.14 / Rs 3,973
753 turns, 273,593,524 cache-read tokens ($136.80 of $177.07 total)
Run /clear between unrelated tasks, or /compact once a session passes ~25 turns. Cost of a long session grows with the square of its length.
200 search/shell calls ran in the main thread (session 7f8ba476)
Est. saving ~$37.38 / Rs 3,290
Grep[path]
Fan out noisy searching to a subagent and keep only its findings. Search output is the highest-volume, lowest-reuse content in most sessions.
Read 13 files, edited none: "[redacted]"
Est. saving ~$5.22 / Rs 459
$10.44 across 44 turns; every file read stays in the main thread's context for the rest of the session
Pure investigation belongs in a subagent - it reads whatever it needs in its own context and returns only the conclusion, so the main thread never pays to carry the file contents forward.
Top-tier model on a 566-token answer: "[redacted]"
Est. saving ~$0.61 / Rs 53
$1.52, model claude-opus-4-7, effort high, 2 turn(s), 566 output tokens
Short, well-scoped answers rarely need the top model or xhigh effort. Drop effort first (it is the bigger lever), then consider Sonnet for this class of task.