Skip to content
Agora
FREN

The footprint of an answer

Under every answer from the assistant, one line shows what that answer used: the tokens processed, the energy, the carbon and the water. We show it for transparency, not to look good. Artificial intelligence consumes energy and water, and a reader is entitled to know it at the moment they use it. The figures are estimates, computed from what the models actually processed for that answer and from published factors, cited at the bottom of this page.

What the four figures mean

Tokens
A token is the unit of text language models read and write, roughly three quarters of a word. The count includes every model that worked on the answer: the one that plans and writes, the three query drafts that check one another, the search of the archive and the final verification. It is therefore far larger than the words you typed or read: most of it is what we give the models to read so that they answer correctly.
Energy
In watt-hours: the electricity of the servers that ran the models, datacenter cooling included. For scale, 3 Wh runs a 250 W desktop computer for about 45 seconds.
Carbon
In grams of CO₂ equivalent, at the average factor of Microsoft’s datacenters based on the grids that actually power them (334 g per kWh). We do not use the “after renewable certificates” factor, close to zero, which says nothing about what was consumed. Our models are deployed globally: a request may be processed in any Microsoft region, not only in Québec, whose grid emits far less.
Water
In millilitres: the water evaporated to cool the servers (0.27 L per kWh) and the water consumed to generate the electricity itself (3.1 L per kWh). The second is the larger share; many publish only the first.

How we compute it

Every model call is recorded with the number of tokens read and written. The energy of a written token depends on the model’s size: 0.0032 mWh per billion active parameters, a factor anchored on measurements of inference servers in production (Oviedo et al., Joule, 2026). A token read counts for 0.1 of a token written: reading is one parallel pass, writing happens token by token. Model sizes are not published by their makers; we take EcoLogits’ estimates. Energy becomes carbon and water through the factors above.

ModelAssumed active parameters (billions)mWh per token writtenmWh per token read
gpt-5.6-sol1950.6240.0624
gpt-5.6-terra980.3140.0314
gpt-5.6-luna390.1250.0125
gpt-4o880.2820.0282
gpt-4.1-mini760.2430.0243
text-embedding-3-small5—0.003

What we do not count

The manufacturing of servers and chips (so-called embodied emissions), the training of the models, the network that carries the answer and your own device. Studies that include them show they add to the figure shown; we prefer a figure we can tie to what the answer actually did. For a method that includes them, see EcoLogits.

Margin of uncertainty

We rent these models; we measure nothing on our own hardware. Publications on energy per token differ from one another by a factor of five to ten, depending on the assumed server utilisation and on what they count. We chose realistic production measurements rather than the most pessimistic ones; take the figures as right within a factor of two. The method carries a version number (2026-10) and every answer keeps the one it was computed with: a future correction does not rewrite what you saw.

Sources