Where your AI effort went

Loading…

Filters

Loading…

Where the effort goes

By repository

Inside

Output by model

Turns by reasoning effort

Tool calls

Named subagents

Heaviest sessions

Writes/reads is the produce-to-explore ratio. Peak context is the largest single-turn context.

What to change

Ordered by severity, then by what it is worth. Computed over the whole window — the filters above do not change it.

highYou are buying tokens at peak rates you did not have to pay≈$21.88/mo

66.9% of your time-priced spend (deepseek, zhipu) ran inside a peak window, where the same tokens cost up to twice the off-peak rate. That timing cost about $39.39 over the period.

  1. Queue unattended work for an off-peak hour — test generation, migrations, doc sweeps, bulk refactors.
  2. Leave interactive work where it is: the premium buys your attention, and a batch job does not need it.
  3. The windows are narrow — DeepSeek 01:00-04:00 and 06:00-10:00 UTC, GLM 14:00-18:00 UTC+8 on weekdays, so weekends are free of it entirely.

Confidence: high

mediumSome sessions carry a very large context per turn

116 session(s) peaked above 300,000 tokens on a single turn (worst: 942,469), and every later turn re-reads it.

  1. Finish the thread and start the next piece of work in a fresh session.
  2. Carry forward only what that work needs — an exploration transcript dragged into unrelated work is what makes context this expensive.

Confidence: medium

lowSessions that paid to cache and then ended≈$0.18/mo

7 session(s) wrote 933,593 tokens to cache and never read one back — the 1.25x write premium paid for nothing.

Ask the follow-ups in the same session — the second turn is where caching starts paying. These were closed right after the first big context load.

Real but immaterial at this volume (~$0.18/month). Kept so you can see the pattern before it grows.

Confidence: high

lowSessions that read a lot and changed nothing≈$0.11/mo

1 session(s) made 7 read-type tool calls with no edit or write, at 0.67 USD. Some is real research; some is hunting for what a targeted search would have found.

  1. Delegate 'where is X' to a search subagent — the answer comes back without the file bodies.
  2. Ask the question you want answered rather than reading toward it.

Real but immaterial at this volume (~$0.11/month). Kept so you can see the pattern before it grows.

Confidence: medium

infoCache is doing its job

89.1% of everything the models read came from cache rather than being re-billed at full input rate.

No action. Watch this number — a fall means sessions are being restarted more often.

Confidence: high

infoSubagent delegation looks proportionate

13.1% of output came from subagents — enough to keep bulk reading out of the main context without runaway fan-out.

No action.

Confidence: medium

infoWhere your AI investment is concentrated

checkout-service took 34.7% of estimated spend across 81 sessions; the top three take 82.1%.

Compare this ranking against where you would say your priorities are. A workspace high on this list and low on your own priority list is the finding.

Confidence: high

infoWhen you actually use AI

Peak hour is 03:00 (379 turns). 42.6% of turns fall outside 06:00-22:00.

Useful as a check on whether AI use is displacing focus time or filling gaps. A high out-of-hours share is worth knowing about for its own sake.

Confidence: high

infoWork is spread across more than one vendor

deepseek is your largest at 36.9% of output tokens, across 5 vendors in total.

No action. A mix is what lets you re-price or re-route without re-learning a workflow.

Confidence: medium