tokensitterAn experimental local hook can read a large cached prefix before it expires, so a later return may avoid rebuilding it.
Local Claude Code plugin · off / experimentalEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
After Claude finishes, a background hook waits, then wakes the session with a one-line note so the cached prefix is read instead of rebuilt after it expires. The tested schedule: A cached prefix of about 100,000 tokens, pinged on the shipped schedule (every 55 minutes, at most 3 hours), then resumed; the cold case is the same resume under a never-seen prefix. It is intended for long-lived caches and large contexts.
The background wake repeats for a limited time after the last user message. The feature is off by default and experimental.
A ping incurs cost. If nobody resumes, there is no saving to offset that cost. Return probability is part of the decision.
Token and cost accounting only; this is not a quality evaluation. A separate check established the short expiry behavior, and the one-hour cold arm used an unseen prefix rather than waiting out the full hour.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Check answers against an objective test, not against how plausible they look.
Include cache writes, cache reads, retries, recalls and helper calls when they apply.
Keep the tested model, task type and cache condition beside the result.
Keep observed misses and untested conditions visible.
Sonnet 5.5 on the Messages API, one-hour cache. 5 repetitions at 100,026 tokens; smaller prefixes of 50,011 and 29,993 tokens ran once each. Resume after expiry, modelled with an unseen prefix and a one-hour cache write, versus a keep-alive ping then a warm resume.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Keep-alive cost study | 5 repetitions at 100,026 tokens; smaller prefixes of 50,011 and 29,993 tokens ran once each 26 Oct 2026 | 90% Cheaper resume after 1 hour idle with keep-alive (Sonnet 5.5, 100,026-token cached prefix; the cold resume is modelled: $0.4004 vs ping + resume $0.0408) +$0.3597 Net saved per resume after 1 hour idle (1 ping; Sonnet 5.5, 100,026-token prefix, cold arm modelled) 85% Cheaper resume after 2 hours idle (2 pings, net +$0.3392) 80% Cheaper resume after 3 hours idle (3 pings, net +$0.3188) $0.0204 Cost of a chain nobody resumes after 1 hour (1 ping), nothing saved 5% Share of idle chains that must end in a resume for the pings to pay off after 1 hour (2 hours: 11%, 3 hours: 16%) | Cache: the one-hour expiry was not waited out; the cold resume is a never-seen prefix, which bills the same full cache write. Cache: how often people come back after an hour or more was not measured; the break-even shares say how often they must. |
|
Lab results at Anthropic API list prices on fixed tasks show what a lever did on those tasks, not what it will do on your bill. For plugin levers on Pro or Max, list-price values describe usage economics, not the subscription invoice. Relay levers apply to API keys only.
Read the measurement method and evidence limitsNo. Lab figures describe fixed tasks and settings. Read the quality outcomes and limits next to each result.
The workload, date, uncertainty and report name are in the table above. The repository is private, so there are no source links, and customer data is never included.
Add Tokensitter to Claude Code and see where your tokens go.