On your device
When you drop an export, a background worker in your browser opens the zip, finds conversations.json, and reads it one conversation at a time. It keeps timestamps, message counts and the length of each message, and throws the text away as it goes. Nothing is uploaded.
We keep a few totals in your browser so your next Wrapped can show your progress. A share link carries only the totals on your card, which our server reads to draw the image.
What we count
Prompts are the messages you sent, including edited and resent ones. Answers are every reply the model generated, including the ones you regenerated. Tool calls, search steps and the reasoning summaries in an export count as hidden steps, and each generated image counts once.
Lengths are estimated at about four characters per token for English and other Latin scripts, and about 1.6 for Chinese, Japanese and similar scripts. Words are tokens × 0.75.
Electricity
Every answer starts from its provider's per-prompt figure: 0.34 Wh for ChatGPT, which is OpenAI's own number, and 0.3 Wh for Claude, which is our estimate because Anthropic publishes none. A typical answer is about 500 tokens; an answer counts for 0.3 + 0.7 × its length ÷ 500 of that figure, so long answers cost more and short ones less.
Reading long inputs costs extra. Epoch AI measured about 2.5 Wh to read 10,000 tokens and about 40 Wh for 100,000; we fit a curve through those points and apply it to the new input of each answer above a typical prompt. Anything over 2,000 tokens is counted as a long paste. Each generated image counts 2.9 Wh.
Water
Water has two parts. Cooling at the data centre uses OpenAI's own ratio for ChatGPT (0.94 mL per Wh) and Amazon's fleet figure for Claude (0.15 mL per Wh). Power plants evaporate about 1.8 mL more for every Wh they generate. Your Wrapped shows both together and the cooling share on its own.
Carbon
Electricity is converted at the average US grid, 390 g of CO₂ per kWh, where most ChatGPT and Claude answers are served. This is what the grid emitted; renewable certificates bought by the providers are not subtracted.
The machine behind it
An AI chip, with its share of servers, networking and cooling, draws about 1.3 kW. Your electricity divided by that is the chip time you used. Your share of the hardware follows from it: 514 kg of CO₂ to make each chip and its server share, 17 kg of hardware per chip, and 27 tonnes of copper per megawatt of data centre, spread over a 5-year chip life and a 15-year building life at 80% use.
Comparisons
OpenAI says ChatGPT gets 2.5 billion prompts a day and has 800 million weekly users, so the average weekly user sends about 3.1 prompts a day. Your prompts per day across your whole period, divided by that, is how many average users you add up to.
Usage levels are ours: light user (fewer than one prompt a day), regular (one to five prompts a day), heavy user (five to twenty prompts a day), very heavy user (twenty to fifty prompts a day), power user (fifty or more prompts a day). Everyday equivalents assume 12 Wh per phone charge, a 10 W LED bulb, 0.17 kWh per km in an electric car, 170 g of CO₂ per km in a petrol car, a 500 mL bottle and a 90,000-word novel. An American home uses 10,791 kWh a year.
What we can't see
Exports leave things out, so your real footprint is higher than your Wrapped. Thinking models write long hidden drafts that mostly aren't in the file. Web searches and agents make extra calls. Training the models isn't counted at all, and neither is your own device or network.
Per-prompt figures are averages or medians. When a provider publishes better ones, we update the Wrapped and the index methodology together.