tokensitterUsage-based lab costs
The lab computes cost from provider usage fields and the pricing table. These are API list-price costs, not reconciled customer invoices.
Features/Savings & measurement
Savings & measurement
Savings you can read.
And reason about.
Read what is observed, what is calculated, and what remains an estimate. Cash cost, plan usage and potential savings answer different questions.
Add to Claude CodeOpenAI Codex: coming soon
Local reporting + API usage evidenceEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Capture usageKeep model, token classes and time window.
Define the comparisonUse a stated baseline and price snapshot.
Show the right unitCash, plan value, or estimate, never an ambiguous total.
Transparent by designEvery claim needs a baseline
The technique
What changes.
And what does not.
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
Counterfactual ledger
The local ledger records actual usage against an original-cost comparison. Some entries have an exact basis; the avoided cost of helper-model calls is still an estimate.
Plan stretch is unverified
Plan value depends on the assumption that plan limits track list-price cost. Plan users (Pro or Max in Claude Code): for plan logins the saving is measured as plan stretch, how much further your plan goes, using plan percentage readings, not dollars. That measurement started on 26 Oct 2026 on one account. No plan figure is published until there are at least 2 weeks of readings.
The boundary. Do not turn helper-model estimates into measured cash or claim a proven plan multiplier. No personal account figures are published.
Quality checks
Fewer tokens is only
half the question.
Ledger reconciliation checks arithmetic and provenance; it does not establish task quality or validate the counterfactual. A helper cost calibration is not a quality test. Optional spot checks are a separate mechanism and are off by default.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Ledger reconciliation
Reconcile task, project and monthly aggregates without duplicate attribution.
Provenance and freshness
Record source, window, rate version and last successful refresh.
Unknown is not zero
Missing or stale records must remain visible as unknown or historical.
Counterfactual validity
Validate the comparison path and disclose which costs are estimated rather than observed.
How we compare
Same task.
A stated baseline.
Each ledger entry requires an explicit original-cost comparison and a basis label. Lab costs use provider token usage and a versioned list-price table; the subscription invoice and the plan allowance are distinct quantities.
Net cash benefit = baseline API cost − optimized API cost − incremental fees
We aim to report alongside: API cash saved · plan usage value · potential savings, separately
We aim to report alongside: API cash saved · plan usage value · potential savings, separately
Figures and evidence
The result.
With its boundaries.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
Source evidenceUsage × ratesLab cost calculation, not a customer invoice
Source evidenceEstimatedAvoided cost of helper-model calls in the local ledger
Source evidenceMeasuringPlan-limit to list-price relationship
Read this alongside the figures. A real call count does not turn an estimated avoided-cost figure into measured savings. A plan-stretch multiplier remains unverified and no plan figure is published.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Plan usage value | One account; readings still being collected. | No publishable plan multiplier until the observation rule is met. | Plan users (Pro or Max in Claude Code): for plan logins the saving is measured as plan stretch, how much further your plan goes, using plan percentage readings, not dollars. That measurement started on 26 Oct 2026 on one account. No plan figure is published until there are at least 2 weeks of readings. |
|
| Ledger classification | Described from the shipped plugin source · snapshot 27 Oct 2026 | Exact-basis and estimated-basis entries are kept separate. | Helper calls are observed; helper savings are estimated. |
|
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 limitsQuestions, answered
Does measured usage mean measured savings?
No. Observed calls can be real while avoided-cost calculations remain estimated. The basis must be attached to the saving, not borrowed from the call count.
Will these pages show a plan-stretch number?
Not until the stated observation period and validity checks are satisfied. API lab costs do not establish a subscription quota exchange rate.
Keep exploring
A little more care.
A clearer picture.
Add Tokensitter to Claude Code and see where your tokens go.
Add to Claude CodeOpenAI Codex: coming soon