Measured August 2026 · fresh paired runs
How much is removed — and does the answer survive?
Two questions matter: how much is saved, and whether the answers stay right. We measured both.
Compression follows your question
CompresLM is a highlighter on a thick binder: it reads your question first and sends the model only the parts that light up.
Pointed questionA little is needed — most of the prompt can go.
Topic questionMore survives, because more is relevant.
"Tell me everything"Almost nothing is safe to drop — so it isn't.
The compression level is not a dial we turn — it follows how much of the document your question actually touches.
Same document, three questions
Only the question changes. Solid is what the model reads and bills; hatched never leaves. (A token is roughly a word.)
Operations handbook — 1,044 tokens, default profile
Fresh run, August 2026. Widths are proportional to real token counts.
Pointed"When does failover to the secondary region trigger?"
−59%Topic"How does the ingest edge handle malformed events?"
−39%Everything"Explain the whole pipeline end to end."
−18%Same staircase on an archive — three medical reports in one prompt:
Archive of three reports — 4,194 tokens, default profile
Ask about one patient and the other two reports mostly disappear.
Pointed"Any abnormal brain findings in Patient B's report?"
−36%Across all"List the abnormal findings for each patient."
−19%Everything"Summarize all three reports, organ by organ."
−5%On a workspace of 70 similar documents, pointed questions removed 42% on average, up to 62%.
The ceiling: when text truly repeats, 98% goes
A worker retrying the same error 300 times, a probe logging the same line all night: one copy answers the question, the rest is pure bill.
Repetitive contexts, default profile — measured, needle checked
In every run the one line that mattered survived — the error, the recovery time, the blocker.
Retry spam — 300 identical errors"What happened and how was it resolved?"
−98.0%Probe spam — 400 identical checks"Did any probe fail? What happened to the pod?"
−98.6%Standup, 30 people, one blocker"What blockers were raised and by whom?"
−76%Text that only looks repetitive (unique timestamps, unique ids) compresses far less — on purpose. 98% is a property of true repetition, not a promise.
Why this is safe
Anchored to your question
Sentences carrying the names, numbers and facts your question is linked to are pinned before anything is scored for removal.
It declines when it must
On a dense contract asked a broad question, the default profile removed 0% — on purpose. When everything matters, nothing goes.
You can always diff it
Send the identical request through a -raw alias and compare. The free tier exists so you can check on your own documents.
Three modes — the only difference is how much they may remove
Not sure which one? Pick auto — it decides per request and you can't get it wrong.
autoRecommended
Reads your question and removes as much as it safely allows — up to 62% on pointed questions, nothing on dense text. The default, and the right choice when in doubt.
fast
The biggest average savings (−28%), aiming at a fixed target. For high-volume traffic where every token counts.
exact
The most careful: trims only true duplicates (−5% on average). For "keep every record" work — lists, invoices, rosters.
Blind-judged, answer against answer
Does the answer survive? Usually — and sometimes it improves
Every measurement is a duel: the same model answers the same question twice — full prompt and compressed — and a stronger model picks the better answer, never knowing which is which.
Two answers
Same question, same model, temperature 0. One prompt whole, one compressed.
Blind judge
A stronger model grades both against the full document. Order shuffled; it never knows which is compressed.
Verdict
Picks a winner or calls it a tie. No partial credit, no self-grading.
All 78 pointed-question duels
Verdicts by the blind judges, both model pairs. The middle is the point: compression usually changes nothing.
When is it better? When there's less to confuse the model
A model reading three near-identical reports can blur them together. A real duel from the run:
"In Patient B's report, there are no abnormal findings in the brain. The MRI shows mild age-related atrophy with mild small vessel ischemic white matter change…"
Contradicts itself: says "no abnormal findings", then lists them. The judge scored it wrong.
"In Patient B's report, there are mild age-related atrophy and mild small vessel ischemic white matter changes noted in the brain."
Names the findings plainly. The blind judge picked this answer — from the smaller prompt.
Same effect on a 125-document archive: the model reading everything mixed up two look-alike reports; the one reading only the relevant pages got it right.
Opt-in: compress_output: true
The answer shrinks too
Half your bill can be the model's own verbosity. Turn on compress_output and the reply gets shorter too — same question, same facts, fewer words:
Length of the model's reply — same question, with and without
Blind-judged: 22 of 24 duels came back equally good either way.
Average across 24 duelsReply without → with
−52%Best case"Walk through what happens to a failed event."
−88%In the best duel the judge even preferred the short reply — "directly and concisely answers what was asked." Those output savings are yours in full: we take no cut of output tokens.
What it's worth
The same percentages, in money
A percentage only means something once it is a line on an invoice. Take the 42% that pointed questions removed across that workspace of 70 documents, and put 100M input tokens a month through it — a middling production workload:
| Model | Bill before | Bill after | You save | Our fee | You keep |
|---|---|---|---|---|---|
| GPT-4o mini · $0.15 / 1M | $15.00 | $8.70 | $6.30 | $1.26 | $5.04 |
| GPT-4o · $2.50 / 1M | $250.00 | $145.00 | $105.00 | $21.00 | $84.00 |
| Claude Sonnet · $3.00 / 1M | $300.00 | $174.00 | $126.00 | $25.20 | $100.80 |
The fee is a share of the saving
Not of your bill, and not a subscription. We take 20% of what compression saved you on input. If a document doesn't compress, you saved nothing and owe nothing.
Shorter answers cost you no extra
With compress_output on, our share of the input saving becomes 50% — and every cent the shorter reply saves on output is yours. We never take a cut of output tokens.
First 1M tokens free
Counted before compression, so the free tier is spent at the rate your bill would have grown, not the reduced one. Enough to measure all of this on your own traffic. Full billing →
Want the arithmetic on your own document rather than ours? The playground prints it — tokens, dollars and our fee — for whatever you paste into it.
Check it on your own documents
Every number above is reproducible on your traffic — the free tier exists for exactly that.