---
id: "lib-2026-11-02-the-nordic-token-tax"
type: "library"
title: "The Nordic Token Tax"
lang: "EN"
org: "Klang"
source_type: "Other"
date: "2026-11-02"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-02-the-nordic-token-tax/"
original_url: "https://research.klang.ai/papers/the-nordic-token-tax"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "A short measurement by the co-founder of Klang, a Nordic speech-AI company, on public parallel datasets; the method is replicable but only ranges are published, and the vendor argues for local labs like itself."
summary_basis: "full-text"
original_minutes: "5"
tags: ["ai technology", "cost", "multilingual", "tokenization"]
---

# The Nordic Token Tax

Tier B (Useful, limited): A short measurement by the co-founder of Klang, a Nordic speech-AI company, on public parallel datasets; the method is replicable but only ranges are published, and the vendor argues for local labs like itself.

Summary based on the full text.

## In short

A 2 November 2026 post by Niklas Silfverström, co-founder of Klang AI. The same 3,000 sentences in English, Swedish, Norwegian, Danish and Finnish were sent through current flagship models and the billed tokens counted. Every model charged at least 30% more for each Nordic language than for English, and up to twice as much for Finnish. Since the Scandinavian translations are the same length as English in characters, the authors attribute the premium to tokenizer vocabularies rather than the languages.

## Key numbers

- +32% to +68%: extra tokens billed for Swedish vs English across frontier models (FLORES-200)
- +56% to +100%: extra tokens for Finnish vs English (FLORES-200)
- 41%: Swedish premium on Mistral Large 4, slightly above the American models
- about 5%: Swedish premium with Viking, an open Nordic model whose tokenizer was built for Nordic text
- 91%, 58%, 35%: Swedish premium of OpenAI's 2020, 2023 and 2024 tokenizers

## What it reports

- Premium ranges on FLORES-200 (encyclopedic text): Swedish +32% to +68%, Norwegian +35% to +66%, Danish +38% to +73%, Finnish +56% to +100%. News text (NTREX-128) gives slightly lower figures, for example Swedish +29% to +64%.
- Models named: GPT-6.1, Claude Opus 5.5, Gemini 4, Grok 4.7, Mistral Large 4, DeepSeek V4, Kimi K3.1, GLM-5.3 and Qwen 3.8, described as twelve models from nine labs.
- Swedish and Norwegian translations match English character length within 0.2%; Danish is 3% longer, Finnish 6%.
- Mistral Large 4 charges 41% extra for Swedish. Its vocabulary has about 50% more entries with French accented letters than OpenAI's, but about as many with å, ä or ö.
- Viking, a Nordic open model from 2024, needs about 5% more tokens for Swedish than English.
- OpenAI tokenizer history for Swedish: 91% (GPT-3, 2020), 58% (GPT-4, 2023), 35% (GPT-4o, 2024); the newest models still use the 2024 tokenizer.
- Japanese premium ranges from 13% to 66%.

## How they got it

Two public parallel-translation datasets (FLORES-200, about 1,000 sentences; NTREX-128, about 2,000) were sent through each lab's newest flagship as of 1 November 2026, and billed token counts recorded. Where a tokenizer is public, counts were also computed locally. The premium is tokens in the target language divided by tokens for the same English text.

## Limits

- Only ranges across models are given; per-model figures are not published apart from Mistral and Viking, and the post names nine models while saying twelve.
- Token premium equals cost premium only for per-token pricing; it ignores price differences between models and output-length effects.
- Klang sells Nordic speech models and the post closes with a case for local AI labs.

## Why it matters for companies

Token budgets and cost forecasts built on English pilots understate spend for staff working in other languages. A Finnish workflow can cost up to twice the English one for the same content, and responses are slower as well. Companies running multilingual assistants should measure tokens per language on their own text before setting budgets or comparing vendors. Where volume is high, a model with a vocabulary built for the language may close most of the gap.

Original: https://research.klang.ai/papers/the-nordic-token-tax