What your chatbot use actually costs
Enter an average day of AI use and this tool adds it up into one carbon and water cost, set against your own footprint and the everyday things around it. Numbers for AI's energy and water come from EcoLogits, an open-source estimator of the energy, carbon, and water behind each AI model. For carbon I use the average grid intensity of the region you pick (not EcoLogits' single whole-world average) plus the hardware's embodied emissions; everything else is sourced in the methodology below. Training is excluded as too uncertain, but split across all users it adds little to your personal footprint. For a fuller carbon picture, and why grid-level changes swamp everything else, see my visual here.
What it writes
"Typical output" is how long each reply is. A "coding / agent session" means one full run of an AI coding assistant that writes and edits code across many steps. It uses EcoLogits' 100,000-token "assist application development" benchmark (much of it the model's own reasoning and tool calls, not final code).
For context, the finished code for this page is about 29,000 tokens, and it took roughly 250,000 tokens of Opus 4.8 output to build it through all the drafts. About the energy of running a dishwasher, or playing a PS5 for 5-6 hours.
·
vs. a day for someone in who lives in , drives , eats , and flies
A day of your AI use vs. everyday things
Excludes training, image and video generation, and retries.
In a year, ways you add emissions
In a year, ways you can cut emissions
How these numbers are made
Per-model AI estimates. Energy, carbon, and water per prompt are computed with the open-source EcoLogits library (v0.10), the engine behind the EcoLogits calculator; the model list is current as of June 2026. EcoLogits estimates electricity from each model's (active) parameter count and output length, and adds the embodied impact of building the hardware. The low to high range is EcoLogits' own 95% confidence interval, mostly uncertainty in the parameter counts of closed models. Water is its water-consumption footprint — water actually consumed (evaporated in data-center cooling and in generating the electricity), not withdrawn and returned. Every water figure on this page is consumptive, so they compare on the same basis.
Carbon basis. EcoLogits gives each prompt's electricity use and its embodied (hardware-manufacturing) emissions separately. We keep EcoLogits' embodied figure as-is and cost the electricity on the grid where you live, using carbon intensities from Ember 2024 and Our World in Data: roughly 380 (US), 215 (EU), 125 (UK), 580 (China), 700 (India), and 480 (world) g CO2e/kWh. Water is left on EcoLogits' global basis.
Output, words, and code. Each row maps to an EcoLogits output-token count: tweet (50), short email (170), article summary (250), chatbot reply (400), 5-page report (5,000), long document (15,000), coding or agent session (100,000), and re-writing the Lord of the Rings trilogy (500,000) — both EcoLogits benchmark tasks; the trilogy itself runs about 480,000 words. The coding/agent session uses EcoLogits' 100,000-token "assist application development" benchmark, which represents one full run that writes and edits code across many steps. Cost scales with length, so long generations dominate. Counts are per day; annual figures multiply by 365. Words assume OpenAI's ~0.75 words per token, read at 238 words a minute (Brysbaert 2019). A coding session is counted as lines of code (~10 tokens a line) rather than words to read. Only about 15% of its output tokens are treated as actual code; the rest is reasoning (which dominates for reasoning models), tool calls, file exploration, and explanation, so a 100,000-token session is on the order of 1,500 lines. This is a rough estimate with wide uncertainty.
Your footprint. "Where you live" is a regional baseline for goods, services, and shared infrastructure (per-capita consumption from Our World in Data); home energy and driving are separate so they do not double-count. Home runs from a small apartment to a big house, about 1.5 to 7 t CO2e a year (EIA RECS; Goldstein et al. 2020). Driving uses EPA's ~400 g CO2/mile across roughly 3,000 to 25,000 miles a year (FHWA). Diet uses food footprints from Poore & Nemecek (2018) and Scarborough et al. (2023): about 1.05 t CO2e a year for a vegan diet up to 3.2 t for a heavy-meat one. Flying is anchored to a transatlantic round trip of about 1.6 t (Wynes & Nicholas 2017); the rarely / sometimes / often options are roughly 0.5, 1.5, and 5 such trips a year. The water footprint uses blue water only — freshwater actually drawn from rivers, lakes, and aquifers. We deliberately leave out green water (rain that falls on farmland and would evaporate anyway, whether or not a crop is grown) and grey water (a notional pollution-dilution volume), because neither is a real draw on freshwater supplies. The AI figure is already blue (data-center cooling and power-plant water), so both sides match. The per-person blue footprint is from the Water Footprint Network (Mekonnen & Hoekstra): about 450 m3/yr for the US and 153 m3 globally (the other regions are scaled by their irrigation intensity and are approximate). Removing green water changes the picture a lot: most of a food's water is rain, so a beef burger drops from ~1,700 L total to about 6 L of freshwater, while irrigated crops stay high.
Lifestyle cuts (the yearly chart). The "saved" figures come from the Founders Pledge Climate & Lifestyle report, drawing on Wynes & Nicholas (2017) (living car-free 2.4 t, avoiding a transatlantic flight 1.6 t, hang-drying, recycling, LED bulbs) and Ivanova et al. (2020) (green electricity, heat-pump heating, a home retrofit, electric and hybrid cars).
Ways you add emissions (the yearly chart). Manufacturing a new car ≈ 6 t CO2e (Berners-Lee, How Bad Are Bananas?); a new flat-screen TV ≈ 350 kg; a new laptop ≈ 250 kg; a new bicycle ≈ 100 kg and a new sofa ≈ 90 kg. Flights, per passenger (myclimate, with a transatlantic round trip ≈ 1.6 t per Wynes & Nicholas): a short-haul round trip ≈ 250 kg, a round-trip US cross-country flight ≈ 1 t, and a transatlantic round trip ≈ 1.6 t. A year of average US driving (~12,000 mi) ≈ 4.8 t at ~0.40 kg/mile (EPA). A beef burger every week for a year ≈ 156 kg (Poore & Nemecek, ~3 kg each); a year of daily coffee ≈ 77 kg; a cotton T-shirt ≈ 7 kg; a new smartphone ≈ 70 kg (Apple); and a pair of jeans ≈ 33 kg (Levi's).
Everyday things. A cup of coffee: ~0.21 kg CO2e and 140 L of water. An hour on a PS5 (~200 W); 3 minutes in a 1,200 W microwave; a mile in a gas car (EPA, ~0.40 kg); a 10-minute hot shower (~21 gal, EPA WaterSense, plus water heating); a dishwasher load (≤3.5 gal, ENERGY STAR); a dryer load (~3 kWh); printing a 400-page book (~2.7 kg, Wells 2012) with paper's water footprint; a beef burger (~3 kg, Poore & Nemecek; ~1,700 L at 15,400 L/kg beef); a new smartphone (~70 kg, Apple) needing ~12,800 L; a pair of jeans (~33 kg, Levi's) at ~10,000 L (Water Footprint Network).
How the appliance figures are derived. The coffee, PS5, microwave, shower, dishwasher, and dryer carbon numbers are each that item's measured electricity (cited above) times the US grid (~0.38 kg CO2/kWh), so they fall on cleaner grids. Two depend heavily on assumptions: a cup of coffee is ~0.05 kg black up to ~0.5 kg for a large latte (we use ~0.2 kg), and a 10-minute hot shower is roughly 0.2 kg on a gas water heater up to 1.5 kg on electric resistance (we use ~0.7 kg). Where a single authority does not exist, figures are rounded mid-range estimates from the sources cited, not false precision.
Water comparisons (blue water). The water charts use the blue (freshwater) component only, so the comparison items are the irrigation- and process-water heavy ones, from the Water Footprint Network and crop studies: a beef burger (~6 L, since beef is ~90% rainwater), a cup of coffee (~4 L), a slice of bread (~8 L), an egg (~15 L), a glass of milk (~22 L), a bowl of rice (~25 L), an avocado (~60 L), a handful of almonds (~120 L, as each California almond takes ~12 L of irrigation), a cotton T-shirt (~1,500 L of its 2,700 L total is blue), a pair of jeans (~5,000 L), and a new smartphone (mostly ultrapure freshwater). The savings come from cutting irrigated products (a few fewer cotton clothes) and outdoor watering: a typical lawn drinks roughly 20,000 gal of freshwater a year, so letting it go unwatered or replacing it with native plants saves most of that (EPA WaterSense). Eating less meat barely moves the blue total, because almost all of meat's water is rain. Energy and manufactured goods have a blue-water cost too: thermoelectric power plants evaporate about 0.47 gal of freshwater per kWh (NREL, cooling/consumption only), so an hour of central air conditioning (~3 kWh) consumes roughly 1.4 gal and a day of the average US home's electricity (~29 kWh) about 14 gal — the very same power-plant cooling water that makes up much of AI's footprint. Printing a 300-page book uses on the order of ~5 gal of process freshwater (paper's water footprint is mostly rain on trees, with the blue mill water lower since most is returned to rivers).
To the extent possible under law, Andrew Masley has waived all copyright and related or neighboring rights to this work. It is dedicated to the public domain under CC0 1.0.