YBacked by Y Combinator

Stop paying for context
your model doesn't need.

Your prompts carry far more tokens than the model actually reads. Compresr drops the rest — up to 90% fewer tokens, so you cut cost and latency. At light compression it matches or beats full-context accuracy on public benchmarks.

$ pip install compresr
$10 free credits · no credit card required
See it on your own file in 60 seconds
▶ Paste into Claude Code
Use compresr to show me live cost savings on my own file.

1. pip install compresr
2. Introspect the SDK to discover the API, don't guess.
3. Ask me for COMPRESR_API_KEY (open https://compresr.ai/signup if I don't have one: $10 free, no card).
4. Ask me for (a) a path to a long document (PDF/.md/.txt) and (b) a question about it.
5. Compress the document with the question, then print a receipt: tokens in/out, ratio, GPT-5.2 cost full vs compressed at $1.75/M input tokens, % saved.
6. Ask GPT-5.2 the question against the compressed context and print the answer.

Don't skip the cost receipt; that's the point.

Works in Claude Code, Cursor, or any agent harness. Open full demo →

~90%
Bill cut
vs. sending the full context
10×
Avg. compression
on long, sparse context
+3.7pp
Accuracy uplift
Compresr 2× vs full context, Pax arena
<40ms
Overhead
p50 API latency at 100K tokens

How it works

We keep the signal and drop the noise.

Feed us the query and the context. We return only the tokens that actually move the answer.

Your raw text112,552 tokens
Boeing 10-K · $0.263 / query
compresr
compressing…
Answer-bearing tokens498 tokens
Compression226×Compressed revenue$77.8B2023 10-Kcommercialairplanes
$0.037 / query · 86% cheaper
Tokens
112,552 → 498
226× fewer
Cost
$0.263 → $0.037
86% cheaper
Latency
18.0s → 13.7s
24% faster

What most teams are doing today

If you're paying full price for your tokens, you're leaving real money on the table.

Trimming

Truncate the tail

Cuts off the end of the context — the answer was often in what you dropped. Accuracy collapses on long docs.

Summarization

Lossy rewrites

Nuance and exact wording are gone. Costs an extra LLM call and adds latency, for a worse context.

Question-agnostic

Compress blindly

Keeps irrelevant tokens, drops important ones. Rarely gets past 5× without tanking accuracy.

With Compresr

One API call. Any scale.

Question-aware compression. Runs in front of any LLM stack — replaces nothing.

Question-aware

Send the query and the context — we keep the spans that carry the answer and drop the rest.

Up to 226× compression

Turn 100K-token prompts into a few hundred tokens. Cut cost and latency without losing the answer.

🎯

Accuracy that holds

At light ~2× compression, accuracy matches or beats full context on public benchmarks.

🧩

Drop-in SDKs

Python and TypeScript clients. Wrap any prompt or document with a single client.compress(...) call.

🔗

Works with your stack

LangChain, LlamaIndex, LiteLLM, and agent harnesses. Compress tool outputs, RAG chunks, or full prompts.

🔒

On-prem ready

Runs inside your VPC. Your data never leaves your network — tuned for regulated workloads.

From prompt to answer

Four steps. One API call.

01

Install the SDK

pip install compresr — or use the TypeScript client, or hit the REST API directly.

02

Send query + context

POST your long document and the question. We support text, PDFs, and tool outputs.

03

Get compressed context

Tokens that carry the answer, nothing else. Typical 10× — up to 226× on long, sparse docs.

04

Forward to your LLM

Send the compressed context to GPT-5.2, Claude, Gemini, or your own model. Pay less, respond faster.

Independent benchmark

FinanceBench — 128 questions over SEC filings.

At light ~2× compression, accuracy holds. Push to ~10× when cost matters more than peak accuracy.

BaselineCompresr
ModelGPT-5.2latte_v2 + GPT-5.2
CompressionNone~2×
Average context~106K tokens~56K tokens
Accuracy73%77%
Cost per queryFull price~47% cheaper

Two ways to deploy

Pick the one that fits your stack.

Hosted SDK

Drop-in SDK. One API key.

Install, grab a key, compress any prompt or document before it hits your LLM. Pay per million tokens, no surprise bills.

  • $10 in free credits on sign-up, no card required
  • TypeScript & Python clients
  • Question-aware compression
  • Transparent per-million-token pricing
Get your free credits →

Sign up, get $10 of compression free, no card needed.

On-prem

Runs inside your VPC.

Your data never leaves your network. We deploy Compresr to your infrastructure, tune it for your workload, and support you directly.

  • Private deployment in your cloud or data center
  • Custom throughput & latency SLAs
  • Tailored to your business needs
  • Dedicated support
Contact us for on-prem →

Enterprise, finance, healthcare, regulated workloads.

FAQ

Frequently asked.

Summarization rewrites the text into something new and lossy — nuance and exact wording are lost, and you pay an extra LLM call. Compresr selects spans of your original text that carry the answer to your query and drops the rest. No rewrite, no hallucinated words.

Any of them. Compresr sits in front of your LLM stack — you send our compressed context to GPT-5.2, Claude, Gemini, Llama, or your own model. We ship first-party integrations for LangChain, LlamaIndex, LiteLLM, and the OpenAI/Anthropic SDKs.

On the hosted API, requests are processed in-memory and never stored or used for training. For regulated workloads, deploy Compresr on-prem inside your VPC — your data never leaves your network.

$0.10 per 1M input tokens. First $10 of credits are free on sign-up — no credit card required. Enterprise on-prem pricing is custom.

At light ~2× compression, our benchmarks show accuracy that matches or slightly beats full-context. Push to ~10× when cost matters more than peak accuracy. You control the ratio per request.

revenue$77.8B token2023 chunkcommercial

Start compressing in 60 seconds.

Install the SDK, grab a key, and see the cost receipt on your own file.

$10 free credits · no credit card required

Contact

Get in touch.

Email us
oussamagabouj@compresr.org
Sales, on-prem, partnerships →
Response time
< 1 business day
Real engineers, not bots.