Two pieces
Tokenization is decided once, at pretraining, and everything downstream inherits it. The obvious fix — train a model with a better tokenizer — costs a full pretraining run, so nobody does it and the inefficiency gets treated as fixed. We change the vocabulary of a model that already exists instead.
Tokens are the unit of billing and of compute alike. A vocabulary that never saw your domain spends more tokens describing it — on every request, for the life of the deployment. The overhead is invisible because there is nothing to compare it against until someone measures.
Windows are counted in tokens. Spending fewer of them on the same material means more real content inside a fixed window — the entire repository rather than a slice of it, the whole trace rather than its tail. This costs nothing extra and is rarely priced at all.
A tokenizer trained on the customer's own traffic encodes that traffic in far fewer tokens than a general-purpose vocabulary can, and the round trip is provably lossless — every byte returns, including the awkward ones.
The saving lands on every request from the day it ships. It does not decay, and it compounds with volume rather than against it.
The same context budget carries proportionally more of your material. Cost falls and capability rises from a single change.
The lexicon belongs to one customer. Without it, the token stream is not readable — opacity is a property of the encoding, not a policy promise layered on top.
Gains vary by domain, so every engagement begins by measuring the customer's own corpus. The number you are quoted is the number from your traffic.
A new vocabulary is of no use if reaching it requires pretraining a model from scratch. The transplant replaces the vocabulary of an already-trained model and re-aligns the model to it, so the model keeps working while speaking a different, private language.
Whitespace-spanning “superword” merging is published work — SuperBPE, COLM 2025 — and we claim none of it. What that research established holds only for models trained from scratch with such a vocabulary, which leaves every already-trained open-weights model unable to benefit from it.
The contribution here is the transplant onto an already-trained model, and the per-domain tuning pipeline built around it.
The companies best placed to build this are the ones billing per token, and halving a customer's token count halves that line of revenue. Their tokenizers are public by necessity, so they can never be the private key to anyone's data. The opening is structural rather than technical, which is what makes it durable.