A drop-in proxy that sends your model only the context it actually needs — for better answers, not just the same ones, at up to 90% lower cost per 1M tokens.
Estimate your own savings
CompresLM sits in front of any LLM API. Before the request leaves, we strip noise and keep the parts that answer the task — then forward the leaner prompt to OpenAI, Claude, or your own model.
Change one line: the base URL. Keep your existing provider key.
Only the tokens that matter reach the model. Nothing else is stored.
Same response shape. Same tools. Lower bill — every single call.
On a paired public-dataset benchmark (200+ cases, same document / same query / same model, run with and without compression), CompresLM matches the quality of uncompressed traffic while cutting the cost of the request — including on medical needle-fact retrieval, a shape of input where a leading internal compression stack showed ~0% savings.
RankAuto v2 is live — up to 90% lower cost on large, unstructured input: sprawling documents, log dumps, and tool-call results that otherwise flood the model's attention.
List price vs. price with CompresLM, same benchmark
Shorter indigo bar = you pay less for the same 1M tokens sent to that model.
Across needle, negation, multi-hop & log-search categories — answer-quality score (0–1)
The two rings are nearly identical — CompresLM answered at least as well, and edged slightly ahead on average, while cost dropped 7–31% on the same cases.
Real per-category results, 200+ paired test cases
Indigo — quality held automatically, this is what you get by default.
Amber — a deeper-compression ceiling; the router only uses it when the safe path can't apply.
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.
CompresLM is a proxy. We process the request, forward it, and return the response. We do not keep your prompts or completions as a product dataset.
First 1 million tokens processed — free. No card required.
After that: pay 20% of what you saved. If we save you $100, you pay $20. You keep $80.
When there is no public token price, we charge a flat $0.50 per 1M tokens saved. Transparent, written out on the billing page.
Measured on this benchmark's compression — actual savings scale with how compressible your traffic is, up to 90%.
| Model | List price / 1M in | With CompresLM |
|---|---|---|
| GPT-4o mini | $0.15 | $0.084 |
| GPT-4o | $2.50 | $1.41 |
| Claude Sonnet | $3.00 | $1.69 |
| Self-hosted / local | $0.50 per 1M tokens saved — no list price to compare against | |
Want on-prem or a custom plan? We help set it up — free. Talk to us.
On-prem setup, a custom plan, or a question — leave a note and we'll get back to you.
Try free. First 1M tokens on us. Then only 20% of what you save.
Create your account