# The Token Company — LLM Token Optimization

> Labs waste your tokens. We don't.

Every filler token in your context is revenue for a lab. With bear-2 in front, filler never wastes the context window. Same answers, a fraction of the cost.

Backed by Y Combinator, and by founders from Hugging Face, Silo AI, Supercell, Dropbox, and Slack.

- [Get an API key](https://thetokencompany.com/meet)
- [Read the docs](https://thetokencompany.com/docs)

## What it does

The Token Company builds proprietary compression models that process raw text before it reaches your LLM. Below 50ms inference, fully deterministic, and cache-safe. One API call compresses your input; you pass the compressed text to any LLM.

Compression is delete-only. A small ML classifier scores every token and removes the ones least likely to affect the answer — nothing is summarized, paraphrased, or generated, so the text that reaches the model is your own, verbatim and in its original order.

## Proven in production

- **Pax Historia — 193B tokens/month.** Compressed prompts outperformed uncompressed in a 268K-vote blind arena across all models: +4.9% Sonnet 4.5, +15% Gemini 3 Flash, +5% purchase lift. [Case study](https://thetokencompany.com/blog/pax-historia)
- **Helonic (YC F25) — construction AI.** Long-running agents analyzing construction drawings save ~47K tokens per prompt. [Case study](https://thetokencompany.com/blog/helonic)

## Built for regulated data

Your prompts are some of your most sensitive data, so we hold them to the same bar you do.

- **SOC 2 Type I** — externally attested. Type II in September 2026.
- **HIPAA compliant** — we sign a BAA.
- **Zero data retention** — optional, set account-wide.
- **On-premise deployment** — your VPC or your data center.

[Visit the Trust Center](https://trust.thetokencompany.com)

## Research

- [Compressing Conversational Context Without Losing the Thread](https://thetokencompany.com/blog/coqa) — CoQA accuracy 93.3% to 95.3% while cutting tokens by 8.2%
- [Introducing Bear-2-Safety](https://thetokencompany.com/blog/bear-2-safety) — cut safety classifier costs by 30% while preserving or improving F1

## Wrap your existing client

One line wraps your OpenAI or Anthropic client. Your existing code stays the same; compression happens automatically.

```bash
pip install the-token-company
```

```python
import requests

response = requests.post(
    "https://api.thetokencompany.com/v1/compress",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "bear-2",
        "input": "Your long text to compress goes here...",
        "compression_settings": {"aggressiveness": 0.1},
    },
)
print(response.json()["output"])
```

[Read the docs](https://thetokencompany.com/docs)

## Frequently asked questions

### What is prompt compression?

Prompt compression, also called **token optimization**, **context compression**, **context optimization**, or **token minimization**, 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.

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

### How does it actually work? Is it generative, or does it just drop tokens?

We sit as a middleware layer between your prompt and your LLM. A small ML classifier scores every token in your input and removes the ones least likely to affect the model's output.

Nothing is summarized, paraphrased, or generated. **We only ever delete.**

That's why we're **faster, cheaper, and 100% deterministic**, where a small-LLM-based compression step is none of those.

### How much will it actually compress my input?

Typically **10–50% while maintaining full accuracy**, depending on how dense your input is.

Clean, information-rich text compresses less. Noisy web scrapes, long chat histories, and verbose documents compress more.

You control how aggressive we are with the aggressiveness parameter.

### How fast is the compression API?

Latency depends on input size, but we're built for real-time use, with **p95 at 150ms**.

For most workflows the shorter prompt after compression cuts time-to-first-token from the downstream LLM by more than the compression step adds, so end-to-end round-trip goes down with compression in the loop, not up.

### Will I lose information I need?

No. The models are trained to remove only the parts of your input that are **not important**: boilerplate, filler, redundant phrasing, structural scaffolding. **Important information is never removed.**

What changes is the **representation**, not the meaning. The text that reaches the model is denser and reads differently, but **the intention of the original is preserved**: the model still sees what you meant.

**We never add information either.** Nothing is invented, paraphrased, or summarized. Compression only ever takes boilerplate away.

### How is this different from summarization?

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

**We only delete.** The remaining text is verbatim, in the original order, which keeps citations, code, numbers, and JSON intact.

### Will it break my LLM provider's prompt caching?

No. Our output is **deterministic** for a given input and setting, so caches (yours or your LLM provider's) stay valid. Cache management is important thus we've designed the compression models to be cache preserving.

### What about data retention and compliance?

**Zero data retention** is available on request and can be set at the account level.

We're **SOC 2 Type I** (audited by Advantage partners, type 2 expected on 09/2026) and **HIPAA-ready** with a BAA.

### Can I run this on-prem or in my own VPC?

Yes. **On-premise deployment** is available on Enterprise: the compression engine runs inside your own VPC or data center, so raw prompts never leave your network.

## Get started

Get an API key and see the savings on your own prompts — [book a time](https://thetokencompany.com/meet) or [contact us](https://thetokencompany.com/contact).