What the evidence says
Two findings, and the second one is the one to read twice.
Icelandic. In the 2025 evaluation, the human translation into Icelandic scored 87.5 and ranked first on its own — no machine system reached the top group. Icelandic is not a large language, but it is comparatively well-resourced, because Iceland digitised it systematically. It is a useful counter-example to the assumption that speaker numbers predict machine translation quality.
Swedish, and the economics. The largest study of post-editing speed ever run — 90 million words, 879 linguists, 11 language pairs, two and a half years of real production — found post-editing into Swedish was 7% slower than translating from scratch. Not less profitable. Slower. Every per-word post-editing discount applied to Swedish in that period destroyed value for whoever accepted it.
And the finding that should change how you read every post-editing average you are quoted: for Polish, the same study found post-editing 18% faster on average — while being slower than translating in 89% of individual jobs. A small number of very fast jobs pulled the mean up. Nearly nine jobs in ten lost time.
Why it happens
Nordic languages are well-resourced, morphologically manageable and structurally close to English. Machine output is good. That is precisely the problem: when the draft is nearly right, the post-editor spends their time reading and deciding rather than typing, and reading takes as long as writing. The saving that post-editing is supposed to deliver comes from not having to compose. In these languages there was not much composing to save.
Finnish is the outlier within the group — agglutinative, structurally unlike English, and harder for engines than its neighbours.
What it means for your workflow
A uniform post-editing discount across a Nordic portfolio is not supported by the best evidence available. And an average is not evidence: the Polish result shows an average can be positive while the typical job loses money.
What we do
We log hours, word count, language pair and content type on every post-editing job we take, from the first one, and we will show you the distribution rather than the mean. Where the data says post-editing is not saving time in a pair, we say so and price it as translation — which is what it is.