Token Screener

$4,536.86 · Rs 399,244
6.15B tokens across 129 sessions · ~$122.62/day
[redacted] · 2026-07-04 → 2026-08-09 · 158 transcripts · 912 tasks

Where it went

Cache read
$2,847.61
6.02B tok · 63% · history re-sent each turn
Cache write
$957.84
100.03M tok · 21% · new context stored
Output
$728.37
30.14M tok · 16% · what Claude wrote
Input
$3.04
656.9k tok · 0% · uncached prompt
Over half of this spend is re-reading context you already paid to create - the cost of long sessions, not of more work.

Daily spend

2026-07-04: $113.952026-07-10: $15.322026-07-11: $287.382026-07-12: $157.612026-07-13: $21.372026-07-14: $230.942026-07-15: $484.982026-07-16: $164.112026-07-17: $23.872026-07-18: $30.222026-07-19: $31.892026-07-20: $156.292026-07-21: $141.142026-07-22: $642.882026-07-23: $536.832026-07-24: $128.332026-07-25: $47.122026-07-26: $135.362026-07-27: $266.542026-07-28: $1.742026-07-29: $0.002026-07-30: $45.272026-07-31: $255.372026-08-02: $14.872026-08-03: $15.652026-08-04: $90.462026-08-05: $204.842026-08-06: $158.332026-08-07: $15.122026-08-08: $7.832026-08-09: $111.252026-07-042026-08-09
Peak day $642.88

By task

TaskCostOutputRe-readTurnsShare
task 001$77.81437.8k116.95M454
task 002$58.34493.7k71.18M254
task 003$52.34405.8k60.50M261
task 004$48.64311.6k70.81M234
task 005$47.7443.9k19.70M37
task 006$45.92313.0k67.03M180
task 007$40.98353.1k53.14M259
task 008$39.52126.0k106.46M249
task 009$39.39310.0k53.93M281
task 010$38.4058.0k64.64M211

By tool (context added by results, estimated)

ToolCallsEst. tokensAvg/callShare
Read156913.63M8.7k
mcp tool 1786310.64M12.3k
Bash50681.16M229
mcp tool 2973591.4k8.1k
WebSearch265203.0k766
ExitPlanMode78200.7k2.6k
Agent156157.1k1.0k
Edit2577130.9k50
Write103848.8k47
mcp tool 1626129.8k114

Model, effort, and thread

SliceTokensCostShare
claude-opus-4-83.89B$3,102.9868%
claude-opus-51.42B$1,075.7924%
claude-sonnet-5832.78M$347.778%
claude-opus-4-77.77M$10.320%
<synthetic>0$0.000%
effort: high2.95B$2,293.6351%
effort: default1.89B$1,499.5433%
effort: medium1.31B$743.7016%
main thread6.15B$4,536.86100%
subagents0$0.000%

Where to optimize

6 findings · ~$1,451.80 / Rs 127,758 estimated avoidable (32% of spend)
561 oversized tool results are sitting in context
Est. saving ~$1,257.19 / Rs 110,633
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.