Published result
What keeps the prompt cache warm (unconfirmed probe)
A plain-language summary comes first, followed by the full text of the result's reproduction pack.
Result summary and pack text · reproduce/api-prompt-cache-reuse-probe-20261104/README.md
- Finding id:
api-prompt-cache-reuse-probe-20261104 - Verdict: Unresolved
- Finished: 4 November 2026
- Comparison: kogen-loop speed, patches and connection routes
- Reproduction pack: reproduce/api-prompt-cache-reuse-probe-20261104/ (data, analysis script and rerun steps)
- Category page: categories/models-and-effort.md
In plain terms
In an 11-call probe, a cached prefix was reused after effort changes and with thinking omitted, but not when thinking was set to between_tools or when the model changed. It is one prefix and one run, so it is an unconfirmed probe and the result is unresolved.
Verdict key: Yes, The result supports the change or claim that was tested. No, The result does not support it, so nothing changes. Unresolved, The evidence is not strong enough either way. Superseded, A later test replaced this result.
Full text of the reproduction pack
The text below is the pack's README as published in the repository (a system folder name may show without its leading slash, so that the site's privacy scan passes).
Verdict: unresolved. This is an unconfirmed probe (n = 11 calls, one prefix, one run), not a confirmed result. In this probe a ~14.9k-token cached prefix (system + two tools + first user turn) was reused after changing effort (low to high and back), after omitting the thinking field, on a mid-conversation per-message effort change (beta), and within one model. It was not reused when thinking was switched to between_tools, nor across models. An earlier run of the same probe showed the same pattern, but no registered confirmation has been run, so the pattern is not stated as a cause or relied on for any decision.
| call | model | effort | thinking | cache write | cache read | result |
|---|---|---|---|---|---|---|
| a | sonnet | low | adaptive | 14901 | 0 | miss |
| b | sonnet | low | adaptive | 0 | 14901 | hit |
| c | sonnet | high | adaptive | 0 | 14901 | hit |
| c2 | sonnet | low | adaptive | 0 | 14901 | hit |
| d1 | sonnet | low | omit | 0 | 14901 | hit |
| d2 | sonnet | low | between_tools | 14900 | 0 | miss |
| e | haiku | low | adaptive | 14901 | 0 | miss |
| h1 | haiku | low | adaptive | 0 | 14901 | hit |
| h2 | haiku | high | adaptive | 0 | 14901 | hit |
| f0 | sonnet | low | adaptive | 0 | 14901 | hit |
| f1 | sonnet | low | adaptive | 0 | 14901 | hit |
11 calls, USD 0.095 at list price (cap USD 5). Each probe run uses a fresh nonce, so no cache is shared between runs. Caveat: one prefix, one run, no repetition; the Haiku rows show the model-keyed cache, not a quality comparison.
Source and runtime
probe.py is a stdlib-HTTP port of an earlier probe script (same calls, same prices).
Rerun
ANTHROPIC_API_KEY=... python3 -I probe.py OUT.jsonl; the table is in data/results.jsonl.