Spending is proportionate.
23.08–22.09: $6,435 at API prices. After the 500K auto-compact, a request fell from 29¢ to 19¢. No background jobs.
median request now · before 16.09: 410K
In money: $6,435
What these requests would cost on the API at Anthropic list prices. A subscription meters limits, not dollars.
On the API that is money. On a subscription it is roughly the same share of your limit: 14%. An estimate from the steps below.
Where the money goes
By model
Money per day
API-equivalent per day. A working day (100+ requests): $379 before 16.09, $178 after.
Table
| Day | Requests | Tokens | API, $ |
| 23.08 | 297 | 125M | $90 |
| 24.08 | 3,100 | 1.52B | $1,026 |
| 25.08 | 2,031 | 960M | $589 |
| 26.08 | 123 | 34M | $30 |
| 27.08 | 1 | 596K | $5.60 |
| 28.08 | 0 | 0 | $0.00 |
| 29.08 | 0 | 0 | $0.00 |
| 30.08 | 0 | 0 | $0.00 |
| 31.08 | 0 | 0 | $0.00 |
| 01.09 | 0 | 0 | $0.00 |
| 02.09 | 0 | 0 | $0.00 |
| 03.09 | 0 | 0 | $0.00 |
| 04.09 | 931 | 307M | $208 |
| 05.09 | 0 | 0 | $0.00 |
| 06.09 | 96 | 57M | $52 |
| 07.09 | 1,919 | 795M | $534 |
| 08.09 | 1,633 | 756M | $495 |
| 09.09 | 1,260 | 531M | $363 |
| 10.09 | 1,200 | 657M | $440 |
| 11.09 | 1,061 | 461M | $351 |
| 12.09 | 804 | 249M | $179 |
| 13.09 | 887 | 334M | $235 |
| 14.09 | 958 | 353M | $273 |
| 15.09 | 2,215 | 593M | $488 |
| 16.09 | 1,043 | 317M | $244 |
| 17.09 | 1,103 | 251M | $206 |
| 18.09 | 1,200 | 289M | $220 |
| 19.09 | 52 | 10M | $12 |
| 20.09 | 108 | 32M | $25 |
| 21.09 | 1,562 | 412M | $323 |
| 22.09 | 514 | 108M | $47 |
The day runs 13:00 to 05:00
93% of requests fall between 13:00 and 05:00 Kyiv time; all 131 sessions came from claude-desktop, no scripts or background agents.
Table
| Hour | Requests | Share |
| 00:00 | 3,441 | 14.3% |
| 01:00 | 2,088 | 8.7% |
| 02:00 | 1,040 | 4.3% |
| 03:00 | 632 | 2.6% |
| 04:00 | 770 | 3.2% |
| 05:00 | 579 | 2.4% |
| 06:00 | 281 | 1.2% |
| 07:00 | 280 | 1.2% |
| 08:00 | 289 | 1.2% |
| 09:00 | 0 | 0.0% |
| 10:00 | 10 | 0.0% |
| 11:00 | 0 | 0.0% |
| 12:00 | 278 | 1.2% |
| 13:00 | 1,108 | 4.6% |
| 14:00 | 941 | 3.9% |
| 15:00 | 1,253 | 5.2% |
| 16:00 | 964 | 4.0% |
| 17:00 | 791 | 3.3% |
| 18:00 | 2,010 | 8.3% |
| 19:00 | 1,309 | 5.4% |
| 20:00 | 1,535 | 6.4% |
| 21:00 | 1,094 | 4.5% |
| 22:00 | 1,095 | 4.5% |
| 23:00 | 2,310 | 9.6% |
Before and after 16.09
Per request, not in total. 'After' is 6 days.
Session start: 78K
33% of the median request. Blue is what you can cut.
CLAUDE.md: 1.1%
All files together over 30 days. Scale goes to 1% of context.
An earlier auto-compact
Replay of 48 sessions (after 16.09). A model, not a measurement.
Saving, % of actual · now 500K
Compactions
What fills the context
Tool output × the number of requests that re-sent it. The largest single file is 0.1% of cache reads.
Setup
What is configured now and what to change.
The median request fell from 410K to 237K. Keep it.
−33% per requestNo plugin adds text to the start of a session. Re-check after plugin updates.
Trimmed on 22.09 from 14.0K to 10.5K characters. The file names in the links are the floor.
≈$115/moThe skill list is 16.8K tokens in every request. Archive the dead ones, unless the list already hits its cap.
≈$55/mo, estimateIts instructions ride in every session. Turn it off and enable it for the session that needs it.
< $5/moWhat to do
Biggest effect first. Together ≈$632/mo, 14% of the current pace of ≈$4,616/mo.
After a break, start fresh≈$300/mo
144 returns after an hour idle rewrote the cache for $615, 10% of the money. Estimate: half is avoidable.
One task, one session≈$277/mo
55% of tokens sit in requests over 300K. Model: a 400K compact saves 6.5% for ~80 compactions a month.
Archive dead skills≈$55/mo
26 of 99 skills unused for 60 days; the list weighs 16.8K tokens. Estimate.
Trim MEMORY.md≈$115/modone 22.09
From 14.0K to 10.5K characters, about 2K tokens off every request.
Not worth it
What the logs can't show
- What counted against the subscription limit: logs show what was sent, not what the server metered.
- Which tasks really needed Opus: 94% of the money is Opus 5, but logs don't say where Sonnet would do.
- Whether the skill list hits its cap: if it does, archiving skills saves no tokens.
- Tool attribution assumes 4 characters per token; for text that undercounts about 2x. Token totals are exact.