German costs AI nearly twice the energy of English. You pay in tokens

You prompt in German, the model answers in German. Feels normal. According to a new University of Michigan study, it isn't: the models tested used on average 1.9 times as much energy for a German answer as for the same task in English. Sounds like a climate footnote. It's not. It's a token bill, and every team automating AI work in German pays it.

What the study measured

Naihao Deng, Rada Mihalcea and colleagues presented “The Language–Energy Divide” at EMNLP; the university published it on November 5, and today it's on the Hacker News front page. They tested three open models: Alibaba's Qwen 3, Google's Gemma 3 and Meta's Llama 3.1. The models answered reading-comprehension questions from the Belebele dataset in 122 languages, plus translated math and everyday prompts, and the team measured energy per response.

English was the cheapest language for every model. Most Western European languages land at twice the energy or less, according to the university; Hindi and Bengali come in at 7.9x and 15.3x. In the published table, German sits at 36.9 joules per response versus 19.5 for English. French is 1.7x, Spanish 2.1x, Italian 2.3x, Czech 4.1x.

The surcharge lives in length, not in the word

Here's the part I find interesting: per token, German needs only about 7 percent more energy than English in the table (0.112 vs 0.105 joules). The real gap comes from the models producing far more tokens for German. My rough math from the table values: about 1.8 times as many per answer. That matches the researchers' explanation: more tokens, more memory, smaller batches, more energy. And because providers usually bill per token, energy turns into money.

For marketing teams in Germany, Austria and Switzerland, that means the surcharge doesn't hit your morning chat. It hits automation. An agent that researches 40 leads, summarizes them, scores them and then drafts an email produces four intermediate stages no customer ever reads. Run every one of them in German and you pay the surcharge four times, for text that only a machine sees. To me this is the unglamorous work before the prompt: a workflow is only as efficient as its process design. AI doesn't fix that, it multiplies it.

The catch

The measurements ran on open models on university hardware, not on ChatGPT, Claude or Gemini. How many tokens German text costs there depends on each tokenizer. The study also finds that high-energy languages often get worse answers. Translate everything into English and you may buy yourself a quality problem. The researchers only recommend translation steps when the translation is reliable.

What I'd do now

• Measure, don't guess: run 20 typical team tasks through your API once in German and once in English and compare the token counts in the response.
• Split the pipeline: an agent's research, classification and internal notes can be English. Anything customers read stays German.
• Keep the final German version in-house: tone and approval belong to the team, not to an automatic translation.

Bottom lineGerman isn't a foreign language to AI, just a pricier one, and the bill lands where nobody is reading.