Private vocabulary for trained models
Inference runs on tokens, and most organizations have never measured what their own traffic actually costs them. We measure it exactly — your corpus, your current tokenizer, on held-out text — then price the same traffic against a vocabulary built for it. Lossless, reproducible, and exclusively yours.
The part nobody prices
The same window, holding more of your content.
A context window is measured in tokens, not in meaning. On a model adapted to the lexicon, the same material in 2.92× fewer tokens means the window you already pay for carries 2.92× more of it — the whole codebase instead of a slice, the full trace instead of the tail. Cost falls and the window holds more, which is why this lands on both sides of a buying conversation.
Shown at the measured 2.916× agentic ratio. A held-out blend of prose, code and agentic traffic in one stream measures 1.97×; tuned to one customer's own traffic it reaches 3.14×.
The line item nobody audits
Which makes the count a cost centre that compounds with every request, for the life of the deployment.
Organizations audit cloud spend, storage and egress. Almost nobody audits the token count itself — the quantity every GPU-second of prefill is spent on. It is treated as a fixed property of the text rather than something measurable, and therefore something reducible.
New engagements
The same discipline — measured on held-out data before anything is committed — applied to three new kinds of work.
One vocabulary covering a language, English and code needs 1.50× fewer tokens than Sarvam-105B’s tokenizer on Hindi, 1.97× fewer than Gemma 3’s on Ukrainian and 1.74× fewer than EuroLLM’s on Polish — on public text, with every held-out document byte-exact.
A model pretrained on the lexicon needs no adaptation step at all, and every training step carries more of your text. The ratio is measured on your held-out text first.
On scanned documents, 2.60× fewer image tokens with no measurable change in answers on two open vision models — measured before anything in your stack changes.
For structured extraction and tool-call work: the baseline and the adapted model on the same held-out tasks, before any production commitment.
Where to look
The first engagement: a lexicon tuned to your corpus, scored on held-out text, so you get your reduction, its dollar value at your rate, and your context headroom before committing to anything.
Token reduction per sector on held-out traffic, the value that represents per terabyte, and the context-window expansion that comes with it.
Lossless round-trip at scale, held-out by content hash, and how much of the gain carries across to tools the lexicon has never seen.
Tell us the domain, the rough volume, and whether you are self-hosted or on an API. That is enough to say what is measurable about it.