tokensitterDefault helper mapping
Explore, researcher, claude-code-guide and the Tokensitter explorer helpers go to Haiku when the caller did not choose a model. This mapping is on by default.
Features/Cheaper helpers
Cheaper helpers
Small tasks.
A smaller helper.
Use a less expensive helper for clearly read-only work while keeping edits and ambiguous tasks out of that path.
Add to Claude CodeOpenAI Codex: coming soon
Local helper hooks · defaults plus opt-in classifierEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Default helper mappingExplore, researcher, claude-code-guide and the Tokensitter explorer helpers default to Haiku when the caller did not choose a model.
Read-only classificationWith
readonlyHaiku on, clearly read-only general-purpose tasks can use Haiku.Quality sampling is separateOptional spot checks rerun sampled tasks on a higher model and score blind.
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.
Read-only classification
With readonlyHaiku on, clearly read-only general-purpose tasks can use Haiku. Edit and ambiguous tasks are not rerouted. Off by default; the recommended setup turns it on.
Quality sampling is separate
Optional spot checks rerun a small sample of finished read-only helper tasks on a higher model and score both blind. They are off by default and use your own API key.
The boundary. The helper study forced Haiku on every easy lookup task; it did not measure the production classifier's accuracy, edit tasks, or real customer work. It does not validate the widget's helper avoided-cost estimate, which remains an estimate.
Quality checks
Fewer tokens is only
half the question.
The answer key is a fixed pattern per question that accepts reference facts and rejects vague answers. Answers correct with the Haiku helper as run against the answer key (Opus helper: 80/80): 77/80. Answers correct, both helpers, after checking the 3 Haiku misses by hand: all 3 were correct answers the key rejected: 80/80.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Task correctness
Check answers against an objective test, not against how plausible they look.
Full cost
Include cache writes, cache reads, retries, recalls and helper calls when they apply.
Workload boundaries
Keep the tested model, task type and cache condition beside the result.
Failure reporting
Keep observed misses and untested conditions visible.
How we compare
Same task.
A stated baseline.
Opus 5.5 main agent at low effort; helper Opus 5.5 or Haiku 5.5. 20 read-only questions about a 16-file, 1,889-line code snapshot: where something is defined, who calls it, how it works, multi-condition traces. The result is full-task cost, not just the helper's own call.
Baseline complete-task cost − feature-on complete-task cost
We aim to report alongside: net API cost per completed task · task pass rate
We aim to report alongside: net API cost per completed task · task pass rate
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.
Lab-70.9%Total billed cost per task, Haiku helper vs Opus helper (95% CI -74.9% to -66.2%)Dilution: the helper is only 71% of the Opus-helper bill in this test, so real sessions with a larger main agent save a smaller share of the whole bill
Lab77/80Answers correct with the Haiku helper as run against the answer key (Opus helper: 80/80)
Lab80/80Answers correct, both helpers, after checking the 3 Haiku misses by hand: all 3 were correct answers the key rejected
Read this alongside the figures. Haiku was forced for all tasks. Production classifier accuracy, edit tasks and hard reasoning were not evaluated. This does not validate the widget's helper avoided-cost estimate.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Haiku vs Opus helper | 160 sessions: 20 tasks x 4 runs x 2 helpers, 80 per helper, none errored 26 Oct 2026 | -70.9% Total billed cost per task, Haiku helper vs Opus helper (95% CI -74.9% to -66.2%) $0.0197 Total billed cost per task with the Haiku helper (Opus helper: $0.0679) -97.5% The helper's own cost, Haiku vs Opus ($0.0483 to $0.0012 per task) 77/80 Answers correct with the Haiku helper as run against the answer key (Opus helper: 80/80) 80/80 Answers correct, both helpers, after checking the 3 Haiku misses by hand: all 3 were correct answers the key rejected | Cheaper helpers: the tasks are easy read-only lookups, so this shows Haiku is enough there, not on hard reasoning or edits. Cheaper helpers: the answer key was widened after the 3 misses, so the as-run score is 77/80 and the audited one 80/80. Cheaper helpers: Haiku was forced on every task; how often the product's classifier picks right was not measured. |
|
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
Is this a production savings guarantee?
No. Lab figures describe fixed tasks and settings. Read the quality outcomes and limits next to each result.
Where can I inspect the evidence?
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.
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