About The Token Company

Our compression models (the current generation is bear-2) learn which tokens in your prompt actually carry the signal the model needs to answer. A small ML classifier scores every token and removes only the low-signal ones, leaving a denser representation of the same information. Nothing is summarized, paraphrased, or generated: what reaches the model is your own text, verbatim and in its original order, just shorter. Because the models only ever delete, output is fully deterministic for a given input and setting, which keeps your prompt caches and your provider's caches valid.

That representation is why accuracy holds flat instead of degrading, and in long-context workloads it can even improve: as a prompt grows, low-signal tokens dilute the model's attention and bury the tokens that carry the answer, so removing them sharpens its focus on what matters. Typical inputs compress by 10-50%, and we have measured accuracy holding flat or rising on FinanceBench, CoQA, and SQuAD 2.0, and with customers running production traffic.

Most customers run compression as middleware: one API call before the LLM call they already make. Python and TypeScript SDKs wrap the OpenAI and Anthropic clients so existing code keeps working. Teams that would rather not touch their request path point their OpenAI SDK at our router instead, which compresses every message before forwarding it to the provider named in the model field.

The workloads that benefit most are the ones with long, repetitive inputs: agent frameworks, RAG pipelines, web research and enrichment, chat apps carrying long histories, and document analysis in legal, financial, and healthcare settings.

Prompts are among the most sensitive data a company has, so we hold them to the same bar our customers do. The Token Company is SOC 2 Type I attested, with Type II expected in September 2026, and HIPAA compliant with a signed BAA available.

Zero data retention can be enabled account-wide on request. Enterprise customers can run the compression engine on-premise, inside their own VPC or data center, so raw prompts never leave their network.

The Token Company is backed by Y Combinator and by founders and operators from Hugging Face, Silo AI, Supercell, Dropbox, Slack, and Wolt.

The Token Company is the trading name of The Token Research Company, a Delaware corporation, based in San Francisco, California.

General enquiries and contract notices go to team@thetokencompany.com. Security reports go to security@thetokencompany.com. See the vulnerability disclosure page for scope and rules. For pricing, an API key, or an enterprise deployment, use the contact form.