Pareto

Pareto has a clean measured fingerprint that currently sits alone: no other measured model shares its signature.

Measured on 5 October 2026 through the endpoint Unbiased serves itself, with a fixed overhead of zero. The prompt counts it reports sit at about half of what any tokenizer on record produces, and on the strings that separate tokenizers they fall to one: a Chinese sentence of 18 characters is billed as one token, a 30-digit number as two, 63 characters of English prose as six. That is close to a word count. No vocabulary produces it, so the reading places nothing and agrees with nothing. The earlier Pareto 26.9 investigation, run while the model was listed as Union Alpha, read real token counts from three tokenizers behind one endpoint; the 26.10 preview, measured the same day as this record, reads as Llama 3 on 48 of 49 strings.

Pareto on OpenRouterPareto 26.10 preview on OpenRouter