Pay only for what we remove.

Prompt compression cuts the number of tokens in an LLM prompt or context while preserving its meaning, so the same request uses fewer tokens and costs less.

Tokeen finds the most token-efficient way to represent your text: every bit of signal stays, the rest goes. It is deterministic and nothing is summarised or rewritten, so what comes back is your text, verbatim and in its original order, minus what the answer does not need.

No. Tokeen sits before your model call, not around it. You make one HTTPS request with the prompt you want shortened, get the compressed prompt back, and pass it to your provider exactly as you do today.

Your provider, your models, your client library, your streaming, your caching: none of it changes. There is no SDK to install and no wrapper to adopt.

You are billed on tokens removed, not on tokens sent. If a prompt cannot be compressed, the call is free. If we remove 20 % of a 10,000-token prompt, you pay for 2,000 tokens, and stop paying your provider for them.

There are no seats, no minimums and no percentage of savings. See the pricing page for the rate.

A small classifier scores every token in the input and removes the ones least likely to change the model's output. Nothing is generated, paraphrased or summarised. We only ever delete.

That is why it is deterministic and cheap, where an LLM-based compression step is neither.

It depends on how dense the input is. Clean, information-rich text compresses less; noisy scrapes, long chat histories and verbose documents compress more.

You control the trade-off with a single rate parameter. Our benchmark shows that keeping 80 % of the tokens is quality-neutral on factual work in French and English; beyond that, test on your own data.

Not by itself. Output is deterministic for a given input and rate, so a prefix you compress once and reuse stays byte-identical across calls, and your provider's cache keeps hitting.

Compress a prefix once, before its first use, and keep it frozen; compress the parts that change on every call.

The model is trained to remove only what does not carry meaning: boilerplate, filler, redundant phrasing, structural scaffolding. Numbers, code, JSON values, URLs and quoted spans are protected.

What changes is the representation, not the meaning. The text reads denser, and the model still sees what you meant. The research page has the measured drop by rate, language and question type.

Summarisation rewrites your input. It changes wording, introduces hallucinations and loses structure.

We only delete. What remains is verbatim and in the original order, which keeps citations, code, numbers and JSON intact.

In Europe. Prompts are processed to produce the compressed version and are not used to train anything.

Security questions before an access request? Ask us.