Thinking AI models emit up to 50 times more CO2 per answer

Every time you ask an AI chatbot a question, it uses electricity. How much depends largely on how much the model has to think before it answers.

That is the finding of a study in the journal Frontiers in Communication, in which researchers measured the CO2 emissions of 14 large language models ranging from 7 to 72 billion parameters. Parameters are the settings a model learns during training, and they are often used as a measure of a model's size. All models were given the same 1,000 questions.

Thinking has a cost

Models with reasoning emitted up to 50 times more CO2 than models that give short answers. Reasoning means the model writes out a series of intermediate steps before giving the final answer. The reasoning models used an average of 543.5 so-called thinking tokens per question. The short-answer models used 37.7. A token is a small chunk of text that the model works with.

The subject matter also plays a role. Complex topics such as abstract algebra and philosophy produced up to 6 times higher emissions than simple topics such as high school history.

The most accurate model was the reasoning model Cogito with 70 billion parameters, which answered correctly in 84.9% of cases. It emitted about three times more CO2 than equally sized models that give short answers.

For comparison, DeepSeek R1 (70 billion parameters) would emit CO2 equivalent to a round-trip flight between London and New York when answering 600,000 questions. Qwen 2.5 (72 billion parameters) could answer about 1.9 million questions with the same emissions.

The researchers themselves point out that the results may depend on hardware, regional power grids and the models studied. That limits how far they can be generalized.

Other measurements and figures from the providers

A media article from The Star about benchmarking company Vals AI's assessment of 16 models reaches similar conclusions. 14 of the models were from Chinese companies. Alibaba's Qwen3.8 Max scored worst on energy, CO2 and water intensity, while Ant Group's Ling 3.0 Flash 2607 scored best. According to the article, long reasoning tasks can have an environmental impact up to 10,000 times greater than simple queries.

Most leading American models from OpenAI and Anthropic are closed and therefore could not be analyzed in the same way.

Google stated in August 2025 that a median text query to Gemini uses 0.24 Wh and emits 0.03 g CO2e. The figure covers only the answer itself and excludes, among other things, training of the model. A text query to ChatGPT is estimated at about 0.3-0.34 Wh, but OpenAI has not shared a detailed methodology. For Mistral, about 1 g CO2 per page of generated text is cited.

According to Hannah Ritchie, who writes about sustainability and data, the providers' figures cannot be compared directly because of differing methods and degrees of openness.