Switching Claude Code’s prompt cache TTL was not worth it either way: on 49 days of my transcripts, a 5-minute cache raises the main thread’s input-side cost 48.3% and a 1-hour cache cuts the subagents’ by 4.3%. Resuming idle workers above 300k context cost more; after a hook started refusing it, that waste fell from 3.5% to 0.38% of subagent spend.
Does this apply to you?
- Claude Code with subagents, used for hours a day: yes.
- Main thread only, no subagents: only the TTL result applies.
- Short sessions without pauses over 5 minutes: the TTL barely matters; a 5-minute cache rarely expires between your turns.
How the replay works
A script parses every transcript and prices each call at API list prices. Then it replays the same calls under a counterfactual:
- The other TTL. A 5–60 minute pause after a 5-minute write turns a cache rebuild into a cache read under 1 hour, and the reverse.
- A fresh worker instead of a resume. An idle worker above 300k context is priced as a new worker with a 100k re-read and a 5k brief.
All dollars are API list-price equivalents, not what a subscription bills. The 49-day replay, 3 September to 21 October 2026, ran in 22 seconds.
Results, 49 days
| Change | Effect | Base |
|---|---|---|
| Main thread 1 h → 5 min | +48.3% | main-thread input-side cost |
| Subagents 5 min → 1 h | −4.3% | subagent input-side cost |
| Fresh worker instead of resuming one above 300k | −8.72% ($59/week) | all subagent spend |
Input-side cost is cache writes, cache reads and uncached input; output tokens are the same under any TTL.
Under a 5-minute TTL, every 5–60 minute pause between main-thread turns rebuilds the cache. Separately, calls above 400k context made 25.8% of main-thread spend and 37.9% of subagent spend over those 49 days, and 12.4% and 12.6% in the latest 14.
What I changed
- Kept the defaults: 1 hour on the main thread, 5 minutes on subagents.
- Added a pre-tool hook that refuses
SendMessageto a subagent above 300k context that has been idle for 60 seconds, so the agent spawns a fresh worker with a brief. A[keep-context: <reason>]tag overrides it. - Made the replay weekly: a timer every Monday 07:30 prices the last 14 days and flags levers over their thresholds, such as TTL ≥5% and resume ≥3% of subagent spend, if they are also worth ≥$10/week.
In the 14 days before the hook merged on 21 October, the resume waste was 3.49% ($19.47/week). In the 14 days to 3 November 06:33 +07, which start one day before the merge, it was 0.38% ($0.63/week). Neither TTL switch is worth it yet: 1 hour on subagents would cut their input-side cost 1.58%; 5 minutes on the main thread would raise its 19.39%.
What went wrong
- The first analysis was wrong. Over the 49 days it overcounted the injected resource-warning text and reported 2.5% of main-thread spend; the replay measures 0.63%. Code review caught it. Cost numbers now come only from the script.
- Usage-limit hits went up. The replay prices dollars, not limits: 33 limit hits in the 14 days to 3 November, against 24 in the 14 days before the hook.
The cost
The replayed spend for those 14 days is $1,887 at list prices: $1,557 main thread, $330 subagents. The weekly check takes 11 seconds.
Check which TTL your Claude Code writes
Sum cache-write tokens by TTL in transcripts changed in the last 14 days. Only jq is needed:
cd ~/.claude/projects
find . -name '*.jsonl' -newermt "$(date -d '14 days ago' +%F)" -print0 | xargs -0 cat \
| jq -r 'select(.message.usage.cache_creation)
| [(if .isSidechain then "subagent" else "main" end),
.message.usage.cache_creation.ephemeral_1h_input_tokens,
.message.usage.cache_creation.ephemeral_5m_input_tokens] | @tsv' \
| awk -F'\t' '{h[$1]+=$2; m[$1]+=$3} END {for (k in h) printf "%-8s 1h=%d 5m=%d\n", k, h[k], m[k]}' | sortMy output on 3 November 2026 06:34 +07:
main 1h=208459501 5m=0
subagent 1h=13968 5m=56169305Main-thread writes are all 1 hour; subagent writes are 99.98% 5 minutes.