Pashto Costs 179 Times English in Energy, a Michigan Study of Three Open Models Finds
A.I. / news
Pashto Costs 179 Times English in Energy, a Michigan Study of Three Open Models Finds
Hindi runs at 7.9 times and Bengali at 15.3 times, across Qwen 3, Gemma 3 and Llama 3.1 on 122 languages.

Generating text in Pashto used about 179 times the energy of English across the same set of requests, according to a University of Michigan study accepted at EMNLP. The researchers measured Hindi at 7.9 times English and Bengali at 15.3 times.
The paper, "The Language-Energy Divide: Measuring Energy Costs of Multilingual LLM Inference," lists eight authors. Naihao Deng, a doctoral student in computer science and engineering, is first author. Rada Mihalcea, the Janice M. Jenkins Collegiate Professor and director of the U-M AI Lab, is corresponding author. Michigan Engineering released a summary on November 5.

1.4 to 180 times, and what was measured
The team tested three open models: Alibaba's Qwen 3, Google's Gemma 3 and Meta's Llama 3.1. The main task was multiple-choice questions from Belebele, a reading-comprehension set covering 122 languages. Energy came from the ML.Energy framework, run on Michigan's Great Lakes cluster.
Two further tasks, grade-school maths and open-ended prompts, exist only in English. The team translated them into seven other languages with Meta's NLLB tool.
Across languages, energy per output token varied by up to 8.3 times. Total energy for a fixed set of requests varied by up to 179 times, from 17.6 kilojoules for English to 3,147 kilojoules for Pashto, according to the paper's abstract page. The release rounds the range to 1.4 to 180 times.
- English1 x
- Hindi7.9 x
- Bengali15.3 x
- Pashto (total energy, 3,147 kJ vs 17.6 kJ)179 x
Source: University of Michigan release via EurekAlert (2026-11-05) and EMNLP paper page, accessed 2026-11-15
Why Hindi and Bengali cost more
Non-English text needs more tokens, the units a model reads and writes. The release says scripts such as Devanagari, used for Hindi, Nepali and Marathi, and Bengali have characters that change shape and combine with accents. That raises token counts.
More tokens take more memory, which leaves less room for processing. The models then work in smaller batches, and energy per token rises. The paper reports that the divide persisted across model sizes, processor types and tasks.
Most Western European languages landed at about twice English or less. The researchers estimate 5.2 billion people speak a first language that costs more than twice as much, using Ethnologue population data.
Cost and accuracy
Many providers charge by token, so a user writing in a high-token language can pay more for the same answer. Google priced its Gemini Decisions API per million input tokens, and NVIDIA's 607.8 tokens a second on two DGX Stations shows how throughput is quoted. Neither figure says what a Pashto token costs.
Languages that used more energy also tended to score lower. Deng said the languages that consume more energy "also tend to have lower accuracy in the model output."
What the study leaves out
Only three open models were tested, so the results may not carry to closed systems. Some benchmarks were machine-translated from English, and the release says translation quality affects results. The 5.2 billion figure combines the measurements with demographic data and is an estimate.
Funding came partly from the National Science Foundation and DARPA, with further support from Ford, the Laude Institute, OpenAI and the Survival and Flourishing Fund. The release is a press summary, and EurekAlert says it is not responsible for the accuracy of releases it posts.
The researchers recommend routing queries to language-specific models, translating prompts to English when translation is reliable, and building more resources for underserved languages. The paper is in the EMNLP 2026 proceedings, pages 48057 to 48075.
Sources
More in A.I.
- 01NVIDIA Nemotron 3.5 Bolt Is Called 4GB; Its 4-Bit File Is 6.18 GBThe Gemma 4 E4B derivative wins seven of eight benchmark rows against its base on NVIDIA's own harness, but loses five of eight to Qwen3.5 4B.
- 02Chrome Built a Fake Web to Test Auto Browse Against Prompt InjectionTwo Chrome AI Security engineers describe a gVisor sandbox that serves generated sites at real URLs, with injections chosen by a request header.
- 03Anthropic Prices IPO at $83, and Two Outlets Disagree on the ProceedsFortune reports about $79.4 billion raised on 980 million shares; The Next Web, citing Bloomberg and the FT, reports at least $68 billion on 801.4 million.
- 04Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says PlannedThe Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.