tokensitterA local prompt instruction
Shorter replies adds one short instruction to each prompt through a plugin hook. It is a plugin feature, not a rewrite of the model's finished answer.
Features/Shorter replies
Shorter replies
Less preamble.
More useful answer.
A short instruction on each prompt asks for more focused output. Ultra is the Claude Code plugin default; Off and Lite give you control.
Add to Claude CodeOpenAI Codex: coming soon
Local plugin · Ultra by defaultEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Choose Off, Lite or UltraThe current plugin defaults to Ultra.
Add a short instructionSkipped when you already have a terse setting.
Evaluate the conversationInclude follow-ups, cost and factual coverage.
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.
Current levels
The shipped instruction has three levels: Off, Lite and Ultra. A former Full level has been retired.
Whole-conversation accounting
The study includes follow-up questions and total billed usage, rather than measuring answer length alone. Cold-session findings should not be read as warm-cache production results.
The boundary. The lab used short synthetic conversation prefixes. The figures come from the written report; the raw per-run file for this study was not saved in the repository.
Quality checks
Fewer tokens is only
half the question.
A blind checklist-based model judge scores required facts across the reply and follow-ups. Judged scores are noisy, so task-specific outcomes are reported instead of a general quality claim. The scripted second-question outcome is separately lower under Ultra.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Requirement coverage
Score every requested item against a fixed checklist.
Executable answers
Compile or run supplied code where the task expects executable output.
Human usefulness
Use blinded reviewers to evaluate clarity and missing context.
Extra-turn penalty
Track clarification requests, corrections and rework.
How we compare
Same task.
A stated baseline.
Total billed API usage across complete conversations, Ultra versus Lite (and Off in the Opus study). Cold sessions with small synthetic prefixes. Warm-cache and added-prefix scenarios in the report are modelled, not measured.
Baseline complete-task output cost − shorter-reply complete-task output cost
We aim to report alongside: output tokens per completed task · completeness score · follow-up turns and failures
We aim to report alongside: output tokens per completed task · completeness score · follow-up turns and failures
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-26%Ultra vs Lite on Opus 5.5, total billed cost per conversation (CI -30 to -21)
Lab97%Facts kept with Ultra after follow-ups (Off: 97%)
Lab93%Facts kept with Ultra on the scripted second question (Off: 98%)
Lab-25%Ultra vs Lite on Sonnet 5.5, total billed cost per conversation (CI -30 to -20)
Read this alongside the figures. Ultra is not equally useful everywhere: the scripted second question scored lower under Ultra than under Off, and a new late question can lose a detail. These figures are copied from the written report, not independently rerun.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Opus 5.5 conversation study | 118 conversations across Off, Lite, Ultra and a former Full level, about 30 per level; a spend cap stopped the second run early 26 Oct 2026 | -26% Ultra vs Lite on Opus 5.5, total billed cost per conversation (CI -30 to -21) -12% Lite vs Off, total billed cost per conversation 97% Facts kept with Ultra after follow-ups (Off: 97%) 93% Facts kept with Ultra on the scripted second question (Off: 98%) | Short replies and lean code ran as cold sessions with a small synthetic cached prefix. Real Claude Code prefixes are larger; with a bigger prefix the percentage is smaller, with a warm cache it is larger. Ultra's weak spot is a new question late in a conversation, where a terse reply can drop a detail. |
|
| Sonnet 5.5 follow-up study | 80 conversations: 2 runs at Lite and at Ultra 26 Oct 2026 | -25% Ultra vs Lite on Sonnet 5.5, total billed cost per conversation (CI -30 to -20) 98% Facts kept with Ultra after follow-ups (Lite: 99%) 86% Facts kept with Ultra on the scripted second question (Lite: 93%) | Do not claim universal parity: the scripted second-question outcome is lower for Ultra. |
|
| Coding tasks | 56 runs: 2 runs of each task with replies at Lite and at Ultra 26 Oct 2026 | -4% Ultra vs Lite on Opus 5.5 coding tasks, total billed cost per task (CI -6 to -2) 28/28 Coding tasks passed with Ultra (Lite: 28/28) | 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
Is Ultra always equally useful?
No universal quality claim is supported. Facts kept with Ultra after follow-ups (Off: 97%): 97%. Facts kept with Ultra on the scripted second question (Off: 98%): 93%.
Does this apply to Codex?
The Codex plugin includes short-reply instructions, but its tests are mock-model only. The lab savings here are not measured Codex savings.
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