Pay for what you compress. Keep what you save.

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

Tokeen finds the most token-efficient way to represent your context: every bit of signal stays, the rest goes. It is fully deterministic and nothing is summarised or rewritten, so the text that reaches the model stays verbatim and in its original order.

We sit as a middleware between your prompt and your provider. 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 faster, cheaper and 100 % deterministic, where an LLM-based compression step is none of those.

Typically 10–50 % depending 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 aggressiveness parameter. Our benchmarks show −20 % is quality-neutral on both French and English prose; above that, check the playground on your own data.

No. Output is deterministic for a given input and setting, so caches, yours or your provider's, stay valid. Stable prefixes stay stable after compression: we never touch the system prompt unless you ask us to, and conversation history is compressed once and then frozen.

Cache reads are where most of an agent's money goes, so being cache-safe by construction is the first thing we measure on every release.

Latency depends on input size, but the compressor is built for real-time use, with a p95 around 150 ms and most calls under 50 ms.

For most workflows the shorter prompt cuts the downstream model's time-to-first-token by more than the compression step adds, so the end-to-end round trip goes down, not up.

The models are trained to remove only what does not carry meaning: boilerplate, filler, redundant phrasing, structural scaffolding. Numbers, code, JSON values, URLs and citations are protected and never removed.

What changes is the representation, not the meaning. The text reads denser, and the model still sees what you meant.

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.

Zero data retention can be enabled account-wide. Prompts are processed in memory and never written to disk. We host in the EU by default.

SOC 2 and a signed BAA are available on Enterprise plans, with the full control list in the trust center.

Yes. On Enterprise the compression engine runs inside your own VPC or data centre, so raw prompts never leave your network. The models are small enough to run on a single GPU node, or on CPU for lower volumes.