Stop paying for words your AI never needed.
Every request you send is billed in full — even the parts the answer never uses. CompresLM removes them on the way through. Same answers, smaller bill, one line to connect.
Works with OpenAI, Anthropic and self-hosted models · on-prem licence available.
What it comes to on your own volume
Put your numbers in
Starts at a measured average — repetitive text goes higher, distinct facts lower.
The idea in one picture
Compression follows your question
Like a highlighter on a thick binder: only what your question needs goes through — and only that is billed.
Pointed questionMost of the prompt can go.
Topic questionMore survives — more is relevant.
"Tell me everything"Almost nothing is safe to drop — so it isn't.
The same effect, measured on real documents:
Pointed"When does failover trigger?"
−59%Topic"How are malformed events handled?"
−39%Everything"Explain the whole pipeline."
−18%A log repeating one error 300×"What happened, and how was it fixed?"
−98%In every measured run, the one line that mattered survived. See how it was measured →
Where your prompt goes — and what we write down
One stateless container. Your prompt is trimmed in memory on ordinary CPU, forwarded to your provider with your key, and freed the moment the call returns. The reply passes straight back to you. The only thing written down is a pair of token counts.
Those two counts are what your savings — and our bill — are computed from. Privacy Policy → · How it works in detail →
Need zero data egress? Run it on your own servers.
The hosted gateway stores nothing — and when policy demands more, CompresLM installs entirely inside your network under a licence. We can set up the LLM itself too, plus data masking — nothing ever leaves your perimeter. How private deployment works →
One line to install
Works exactly like calling OpenAI, Claude, or a local LLM directly.
# Python — OpenAI SDK
client = OpenAI(
base_url="https://compreslm.com/v1",
api_key="YOUR_COMPRESLM_TOKEN",
default_headers={"X-LLM-API-Key": "YOUR_PROVIDER_KEY"},
)
We never store your provider key. It rides with each request and is discarded after the call.
Or let your coding agent do it
Copy this into Claude Code, Cursor, or whatever you already have open. It walks your agent through the wiring, stops to ask you for the two things it cannot do itself, and finishes by running the same question with and without compression so you see the saving and the answer side by side — on your own material, not ours.
Pricing that can only save you money
The first 1,000,000 tokens (≈750,000 words) are free, no card required. After that you pay 20% of the money we actually save you — a bill that can only exist if the saving does.
What you get
Six things that start the moment you change the base URL.
The biggest prompts shrink most
Logs, tool output, boilerplate — the more it repeats, the less is sent.
Nothing is invented
We leave words out. We never write new ones in.
Sometimes the answer improves
In blind duels the shorter prompt beat the full one — less noise to distract the model.
Send the whole archive
Pay only for the part your question actually needs.
Replies shrink too
And output costs several times input — reasoning models bill their thinking as output.
Your text stays yours
No database, no cache, no training, no third parties.
Need it fitted to your setup?
Closed network, unusual documents, a hand getting it running — tell us and we will help.
The questions everyone asks, answered plainly
Frequently asked questions
What is CompresLM and how does it work?
A proxy between your app and your LLM provider: you keep your provider, model and key, and change one line — the base URL. Before each request is forwarded, it removes the parts of the context your question never needed — like an assistant crossing out the 24 pages of a 30-page contract that don't concern your question. Fewer tokens reach the provider, so the provider bills you less. Measured: −59% of a prompt on a pointed question, answers blind-judged same or better in 95% of pointed cases. How it works →
What makes it better than the alternatives?
Four things, measured rather than promised. Compression follows your question — sentences carrying the names and figures you asked about cannot be dropped. Nothing is rewritten — words are removed, never invented. Nothing is stored — your text passes through memory and is gone. And the fee is 20% of the measured saving, so on known-price models it cannot cost more than it saves. Setup stays one base-URL change, on any provider. See the measurements →
How is this different from RAG, prompt caching or LLMLingua?
They cut different parts of the bill, and CompresLM composes with all of them. RAG fetches folders from your archive — CompresLM crosses out the irrelevant pages inside whatever was fetched, so you can over-fetch and let compression pay for it. Provider prompt caching discounts a repeated prefix from the second call on — compression removes tokens before they are billed at all, on every call. Libraries like LLMLingua do compression inside your code, hosted and tuned by you; CompresLM is that idea as a hosted proxy: no code changes, no models to run, public quality benchmarks. More in the docs →
Is my data stored anywhere?
No. Think of a water meter: it counts litres but keeps no water. Your prompt is compressed in memory, forwarded to your provider on your own key, and freed the moment the call returns; the reply is handed straight back and never cached. What is written down is numbers — token counts, costs, timestamps. Provider keys are never stored either. Privacy Policy →
Can I run it on my own servers?
Yes. The hosted service already stores nothing — and when policy forbids any third-party processor at all, CompresLM can be installed entirely inside your own network, on your VPC or on-premise servers, under a licence agreement. We can set up the LLM itself too — an open-weights model with tuned inference — plus data masking for the traffic, so nothing leaves your perimeter at all. Terms are individual and we help with the setup. How private deployment works →
What does it cost? Is there a subscription?
There is no subscription and no minimum — you pay only out of what you save. The first 1,000,000 tokens processed are free (2,000,000 once a card is on file; adding it takes no payment). After that the fee is 20% of the dollars we measurably saved you — 50% of input savings when reply compression is on, or $0.50 per 1M tokens saved on local models. Charged when the balance reaches $50, or monthly. If we save you nothing, you pay nothing. How the bill is calculated →
How does the referral program work?
Share your link from the dashboard's Referrals tab and you earn 10% of every payment a referred account makes during its first 6 months — one flat rate, no tiers to climb. A commission becomes withdrawable 60 days after the payment (the refund window) and pays out to Payoneer once your balance reaches $50. Nothing accrues on free-tier usage. Full program terms →
A different question? Ask the assistant — it answers in your language — or write to us.
Talk to us
- Connect it — we'll get it running against your traffic and check the numbers with you.
- Fit it to you — closed network, unusual documents, a shape we don't handle yet.
- Build it — no product to connect yet? Our partners build it, we make it cheap to run.
Usually a reply the same working day.
Start saving today
First million tokens free. No card, no sales call.
Create your account