The same AI question costs three times the energy in Estonian as in English, a new study measures

Researchers from the universities of Michigan, Aberdeen and Cambridge measured how much electricity a language model uses to answer the same questions in 122 languages. Their paper, presented at the EMNLP 2026 conference in Budapest, puts Estonian at about three times the energy of English. ERR Novaator picked up the result on 10 November.

The test used Qwen3-8B, a small open model, on a single Nvidia L40S graphics card. Each language got the same 900 reading comprehension questions from the Belebele benchmark. English answers used 16,762 joules in total. Estonian answers used 51,689 joules, roughly 3.1 times as much. The widest gap in the study was between English and Pashto, at 179 times.

Most of the extra cost is longer answers

Each token of Estonian output cost slightly more energy than English, 0.12 joules against 0.10. The bigger factor was length. The model wrote 430,745 tokens of Estonian against 167,621 of English for the same questions, about 2.6 times as many. The authors explain that tokenizers trained mostly on English split other languages into more pieces, and that models write longer when they understand a language less well.

The longer Estonian answers were also less often right. The model scored 76.2% in Estonian and 94.6% in English. Our neighbours fared similarly: Finnish and Lithuanian came out at about 2.8 times English, Latvian at about 4.5 times. German and French stayed under twice English, and Spanish came in just over.

What the study does not show

The paper measured open models running on research hardware, not ChatGPT, Copilot or Gemini. Commercial services use larger models and their own tokenizers, so the Estonian factor there could be smaller or larger. The authors did repeat parts of the test on four other models and two kinds of GPU, and the gap between languages held every time, though Estonian was not in that smaller set.

The money link is the part an Estonian company should care about. Most AI APIs bill per token, so a task that produces 2.6 times the tokens produces a bigger bill. The authors also warn that because energy drives inference prices, providers may end up charging speakers of smaller languages more, or dropping those languages.

Their main suggestion for builders: for tasks where the language does not matter, translate the input to English, do the work in English and translate the result back. They report this can cut energy and often improves accuracy. They also warn against it for anything that depends on tone, idiom or cultural context.

What to do on Monday

• ✳Open the usage page of the AI API you pay for and find your three most frequent tasks. Run each once in Estonian and once in English and compare the token counts
• ✳Set a length limit in your prompts, such as "answer in three sentences". In this study the extra cost came mainly from longer answers
• ✳For internal, language-neutral work such as pulling fields out of invoices or sorting support tickets, test whether English instructions give the same result for fewer tokens
• ✳Keep Estonian for anything a customer reads and anything that depends on tone. The authors themselves warn against translating those tasks
• ✳Before you pick a small or self-hosted model, test it on 20 real Estonian examples from your own work. A score in English says little about Estonian
• ✳When you compare AI vendors, ask for the price of a typical Estonian task, not the price per million tokens

Nobody has to stop using Estonian with AI. But if your AI bill is growing, check how much of it is your language before you blame the model.