tokensitterA short instruction
The lean-code instruction (under 60 tokens) is added through the plugin hook. It asks Claude to plan first and change only what the task needs.
Features/Lean code
Lean code
Smaller changes.
Less unnecessary code.
A short local instruction asks the coding agent to plan first and change only what the task needs. Lite is the plugin default.
Add to Claude CodeOpenAI Codex: coming soon
Local plugin · Lite by defaultEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Set the instructionOff, Lite or Full; Lite by default.
Plan before editingFocus the change on the requested task.
Run the task checksHidden tests and visible tests judge the result.
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.
Three levels
Off, Lite and Full were compared in the small coding-task study. On these tasks a stronger setting did not save more: Full saved less than Lite.
Judge the code, not the size
A smaller diff is not sufficient. The study checks hidden tests and that visible tests still pass; token and billing comparisons are reported separately.
The boundary. Results cover small tasks only: 14 small TypeScript fixes and features with hidden tests, run by a headless agent with 5 tools. Passing every sampled run does not establish unchanged quality on large refactors, other languages, or real customer work.
Quality checks
Fewer tokens is only
half the question.
Pass means all hidden tests pass and no visible test fails. Tasks passed at every level: 28/28. That small sample leaves uncertainty. Lower end of the 95% Wilson interval for the passes: 88%. Zero observed failures does not prove zero regression risk.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Instruction preservation
Check whether constraints and user requirements survive the reduction.
Reference completeness
Verify necessary paths, errors, dependencies and source details remain available.
Long-session outcomes
Test multi-step tasks where earlier context becomes relevant later.
Cache and rework
Include cache misses, repeated reads and additional calls.
How we compare
Same task.
A stated baseline.
Lean code Off versus Lite, total task cost, separately for Opus 5.5 and Sonnet 5.5. The study also tested Full. 168 runs: 14 tasks at Off, Lite and Full, 2 runs each, on both models.
Baseline full-session cost − optimized full-session cost
We aim to report alongside: input and cache tokens by class · completed-task cost · regression and rework rate
We aim to report alongside: input and cache tokens by class · completed-task cost · regression and rework 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-36%Lite vs Off on Opus 5.5, total billed cost per task (CI -39 to -33)
Lab-14%Lite vs Off on Sonnet 5.5, total billed cost per task (CI -19 to -9)
Lab28/28Tasks passed at every level
Read this alongside the figures. The raw per-run file for this study was not saved in the repository. Do not generalize small TypeScript tasks to large refactors, or claim unchanged quality.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Lean-code task study | 168 runs: 14 tasks at Off, Lite and Full, 2 runs each, on both models 26 Oct 2026 | -36% Lite vs Off on Opus 5.5, total billed cost per task (CI -39 to -33) -26% Full vs Off, total billed cost per task -14% Lite vs Off on Sonnet 5.5, total billed cost per task (CI -19 to -9) -13% Full vs Off, total billed cost per task 28/28 Tasks passed at every level 88% Lower end of the 95% Wilson interval for the passes | The coding tasks are small: every level passed every task, so a quality cost would not show at this size. |
|
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
Why is Lite the default?
On the tested small tasks, Lite reduced Opus cost more than Full. This is a scoped result, not a reason to assume every workload behaves the same.
Does this compress input context?
No. Lean code is an output and behavior instruction. Relay compression, Read less and reversible context are separate mechanisms.
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