KN-04 · Cache and compaction economics
When an agent session keeps adding context, each call can send more for the model to read. A prompt cache lowers the price of rereading shared content. Compaction can reduce the amount sent.
If each call adds the same amount of retained context and later calls reread it all, cumulative reads grow roughly with the square of the number of calls. Real sessions can prune content, compact it, or add uneven amounts. Cache pricing changes the cost per reread. The accumulation pattern depends on the session.
The study used these historical Claude API list rates in USD per million tokens.
| Model | Input | One-hour cache write | Cache read |
|---|---|---|---|
| Opus 5 | $5 | $10 | $0.50 |
| Opus 5.5 | $4 | $8 | $0.20 |
A cache read dropped from 10% to 5% of the input rate. Meanwhile, the ratio of a one-hour write to a read rose from 20 to 40. In this comparison, warm reads got cheaper, while a one-hour rewrite cost twice as much relative to a read.
The study traced 313 calls across nine sessions. In one session, compaction took the context sent per call from 382,000 tokens to 38,000. It grew back to 172,000 over the next 11 calls. The refill included 12 web fetches, two searches, one file read, and the model’s own long output sent back as input.
The app’s savings display compared the old context with the summary alone. The next request also carried a fixed prefix. In three compactions the actual drop was about 29,000 tokens smaller than the display implied. In a separate set of 196 Codex compactions, the median context reduction was 87%.
Figure 01Field measurement
Three observed points from one session among nine traced sessions and 313 calls. The intermediate calls are outside the chart. The next request includes both the summary and its fixed prefix.
On one agent seat, manual compactions happened earlier than automatic ones, and calls read less cached context. On a second seat, manual compactions came later and calls read more. These two cases suggest looking at timing before choosing a trigger.
Compaction also interrupted work. The waits were 9 to 57 seconds in one day of Claude Code compactions, and about one to three minutes in nine Codex compactions across three chats. Frequent compaction can interrupt an active task and remove context it still needs.
At the listed Opus 5.5 rates, reading 100,000 cached tokens once costs about US$0.02. Writing all 100,000 back into a one-hour cache costs about US$0.80. The read-only arithmetic reaches the cost of one rewrite after 40 reads.
This comparison assumes each read hits and refreshes the same cached prefix in time, with no extra input or output charged. Actual keep-alive calls can add cost. A fresh session with a short handoff can cost less to start, but the agent may need to recover missing details before it can work.
Figure 02Evidence comparison
Quantity100,000 cached tokens
Estimated API costUS$0.02
Quantity4 million cached tokens in total
Estimated API costUS$0.80
Quantity100,000 tokens in a one-hour cache
Estimated API costUS$0.80
Illustrative arithmetic for 100,000 tokens at the historical Opus 5.5 rates used in this study. Each read assumes a hit on the same prefix and excludes extra input and output. Subscription usage and recovery work are outside the estimate.
The nine sessions came from one seat in late September 2026. The refill example came from one session, and the trigger comparison used two seats. The team recorded the API prices in October 2026. The idle example is arithmetic on those rates. It does not show how subscriptions count cache use. We did not rerun the measurements for this publication.
Related field notes: What your agent still knows after a compaction and Where the tokens go in a real agent team.