The Token Company: LLM Gateways

The Token Company builds a drop-in compression API that preprocesses LLM inputs using fast ML models to remove redundant tokens from prompts, chat histories, and RAG documents. Part of YC W2026, it was founded by Otso Veistera — reportedly the youngest solo founder in YC history at 18 years old. YC partners reached out directly rather than through the standard application process.

The API compresses 100K tokens in under 100ms using purpose-built classification models (bear-1 series) that identify and strip low-value tokens. This is not a generative LLM but a fast, deterministic model. The counterintuitive result: compression actually improves accuracy because models focus on higher-signal content. Published benchmarks show +2.7pp on financial QA with 20% fewer tokens and +4.0pp on reading comprehension with 17% fewer tokens.

A named customer, Pax Historia (processing 193B tokens/month), ran a 268K-vote blind arena study showing compressed prompts outperformed uncompressed with a +5% purchase lift. The pricing is straightforward at /bin/zsh.05 per 1M compressed tokens, and you only pay for tokens actually removed.