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.

237K

median request now · before 16.09: 410K

9.15B
tokens in 30 days
98%
of them cache reads
24,098
requests, deduplicated
131
sessions, all from claude-desktop

In money: $6,435

What these requests would cost on the API at Anthropic list prices. A subscription meters limits, not dollars.

≈$154
per day at the current pace
≈$4,616
per month at the current pace
≈$632
per month you can save · 14% of spend
29¢ → 19¢
per request before and after 16.09 · −33%

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

Cache reads · 68%Cache writes · 21%Output · 11%

By model

Opus 5$6,051
Fable 5.1$302
Sonnet 5$45
Opus 5.5$28
Opus 4.8$8.55
Haiku 4.5$0.20

Money per day

API-equivalent per day. A working day (100+ requests): $379 before 16.09, $178 after.

before 16.09after
23.08
30.08
06.09
16.09
22.09
Table
DayRequestsTokensAPI, $
23.08297125M$90
24.083,1001.52B$1,026
25.082,031960M$589
26.0812334M$30
27.081596K$5.60
28.0800$0.00
29.0800$0.00
30.0800$0.00
31.0800$0.00
01.0900$0.00
02.0900$0.00
03.0900$0.00
04.09931307M$208
05.0900$0.00
06.099657M$52
07.091,919795M$534
08.091,633756M$495
09.091,260531M$363
10.091,200657M$440
11.091,061461M$351
12.09804249M$179
13.09887334M$235
14.09958353M$273
15.092,215593M$488
16.091,043317M$244
17.091,103251M$206
18.091,200289M$220
19.095210M$12
20.0910832M$25
21.091,562412M$323
22.09514108M$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.

13:00–05:00other
00
03
06
09
12
15
18
21
Table
HourRequestsShare
00:003,44114.3%
01:002,0888.7%
02:001,0404.3%
03:006322.6%
04:007703.2%
05:005792.4%
06:002811.2%
07:002801.2%
08:002891.2%
09:0000.0%
10:00100.0%
11:0000.0%
12:002781.2%
13:001,1084.6%
14:009413.9%
15:001,2535.2%
16:009644.0%
17:007913.3%
18:002,0108.3%
19:001,3095.4%
20:001,5356.4%
21:001,0944.5%
22:001,0954.5%
23:002,3109.6%

Before and after 16.09

Per request, not in total. 'After' is 6 days.

before 16.09after
Median request−42%
410K
237K
p90 request−42%
784K
451K
Cache reads per day−53%
456M
216M
Tokens in requests >600K−78%
43.8%
9.6%
Session peak, median−45%
466K
255K
Price per request, API−33%
28.9¢
19.3¢

Session start: 78K

33% of the median request. Blue is what you can cut.

start 78Kconversation 158K
tool schemas (remainder)27.8K
skill list16.8K
memory index MEMORY.md9.0K
built-in system prompt7.3K
deferred tool names6.5K
MCP server instructions2.4K
subagent list2.1K
email + git status1.9K
project CLAUDE.md1.3K
global CLAUDE.md1.0K
small bits (env, date, model...)0.7K
first user message0.3K

CLAUDE.md: 1.1%

All files together over 30 days. Scale goes to 1% of context.

project-a0.39%
project-b0.23%
project-c0.15%
project-d0.15%
project-e0.10%
project-f0.06%
project-g0.04%
project-h0.03%
project-i0.00%
project-j0.00%

An earlier auto-compact

Replay of 48 sessions (after 16.09). A model, not a measurement.

Saving, % of actual · now 500K

10.9%300K
6.5%400K
0.5%500K
0.0%600K
0.0%750K

Compactions

24300K
16400K
1500K
0600K
0750K

What fills the context

Tool output × the number of requests that re-sent it. The largest single file is 0.1% of cache reads.

Bash797M
Read765M
Claude_Browser · browser_batch196M
Claude_Browser · javascript_tool71M
claude-in-chrome · computer61M
connector · read_file_content53M
connector · jobs_wait52M
n8n-mcp · n8n_list_workflows51M
claude-in-chrome · browser_batch47M

Setup

What is configured now and what to change.

✓ finesettings
Auto-compact at 500K

The median request fell from 410K to 237K. Keep it.

−33% per request
✓ finehooks
No start-up hooks

No plugin adds text to the start of a session. Re-check after plugin updates.

✓ finememory
MEMORY.md: 5.9K tokens

Trimmed on 22.09 from 14.0K to 10.5K characters. The file names in the links are the floor.

≈$115/mo
i ideaskills
26 skills unused for 60 days

The skill list is 16.8K tokens in every request. Archive the dead ones, unless the list already hits its cap.

≈$55/mo, estimate
i ideamcp
An MCP server used 3 times a month

Its instructions ride in every session. Turn it off and enable it for the session that needs it.

< $5/mo

What to do

Biggest effect first. Together ≈$632/mo, 14% of the current pace of ≈$4,616/mo.

$300

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.

−6%

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.

$55

Archive dead skills≈$55/mo

26 of 99 skills unused for 60 days; the list weighs 16.8K tokens. Estimate.

−2.5%

Trim MEMORY.md≈$115/modone 22.09

From 14.0K to 10.5K characters, about 2K tokens off every request.

Not worth it

Splitting CLAUDE.mdall 11 files together are 1.2% of context
Auto-compact at 300Kmodel: −11% for ~120 compactions a month
Hunting a single file or toolthe heaviest file is 0.1% of cache reads; Bash and Read ~9% each, no single culprit

What the logs can't show