tokensitterDeterministic trimming
Long tool results retain a head and tail; repeated lines collapse and ANSI formatting is stripped. The ruleset is versioned.
Features/Token compression
Token compression
Less tool noise.
A steadier context.
Deterministic rules shrink new tool results at the API relay. No extra model call is needed for this transformation.
Add to Claude CodeOpenAI Codex: coming soon
API relay · opt-in per keyEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Deterministic trimmingLong tool results retain a head and tail; repeated lines collapse and ANSI formatting is stripped.
Byte-stable prefixesThe transformation is designed to be idempotent so repeated application does not keep rewriting a cache prefix.
Off until enabledRequest compression is opt-in per API key.
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.
Byte-stable prefixes
The transformation is designed to be idempotent so repeated application does not keep rewriting a cache prefix. Safe and lossy modes remain explicit choices.
Off until enabled
Request compression is opt-in per API key. It is separate from plugin Read less and from reversible storage with recall.
The boundary. The long-run result combines routing, compression and Read less; it is not a compression-only effect. The confidence interval includes zero. Whether the lever is live on the production relay has not been verified.
Quality checks
Fewer tokens is only
half the question.
The combined agent benchmark uses objective task checks, including test files restored before execution. It does not isolate compression, so quality outcomes cannot be attributed to compression alone. The pass checks were hardened after the runs, and the stored run records were not re-graded under the hardened checks, so the pass counts are as graded at the time.
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.
Prompt-cached Opus 5.5 versus an arm that also enables routing, cache, compression and Read less. The synthetic direct-API harness imports the relay logic; it does not test the live relay service.
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.
Source evidenceDeterministicRule-based tool-output shrinking
Lab · No proven saving-6.5%Long agent runs, parallel tool calls allowed: routing, cache, compress and Read less together vs a prompt-cached Opus agent, total billed cost per run (CI -19.4 to +2.6, includes 0, 15 pairs, passed 15/15 both ways)
Source evidenceNot isolatedNo compression-only saving established here
Read this alongside the figures. No standalone compression percentage is supported by this comparison. The confidence interval includes zero, and the combined intervention cannot identify an individual feature effect.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Combined long-run arm | Arm: routing, cache, compression and Read less together, run with parallel tool calls allowed. The serial-run comparison did not include this arm. 15 pairs per comparison: 5 tasks x 3 repetitions x 2 arms (cache baseline, routing per step plus cache), run twice (one tool call per turn, and parallel tool calls allowed); all passed in both arms 27 Oct 2026 | -6.5% Long agent runs, parallel tool calls allowed: routing, cache, compress and Read less together vs a prompt-cached Opus agent, total billed cost per run (CI -19.4 to +2.6, includes 0, 15 pairs, passed 15/15 both ways) No proven saving | Route, long runs: only 5 synthetic coding tasks (15 pairs per comparison, one date). The intervals are wide, so a small saving or a small extra cost cannot be ruled out. |
|
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