The itemised bill

01

Writing the answer

human, on a keyboard

52 words/min

fastest AI

1,397 tokens/s

× 1,200

A token is about three quarters of a word: 52 words a minute is 1.2 tokens a second. The smartest models (GPT-6.1 Sol, Claude Opus 5.5) write 53 to 96 tokens a second, 46 to 83 times a human.

Sources : Aalto University, 2018, 168,000 volunteers · Artificial Analysis, read on 9 Oct 2026 · OpenAI, the three-quarters rule

02

Waiting for the answer

human, by email

47 min

AI, on this site

Half of all email replies leave within 47 minutes, across 16 billion emails studied. The other figure is measured here, over the answers this site has written. A frontier model set to maximum reasoning takes 25 seconds to 11 minutes before its first word.

Sources : Kooti et al., WWW 2015, 2 million users · Artificial Analysis, time to first word, read on 9 Oct 2026

03

Reading the answer

you, reading

238 words/min

frontier AI

53 to 96 tokens/s

× 10 to × 18

238 words a minute is 5.3 tokens a second. The answer is on screen before you finish its first line: the slow one is you.

Sources : Brysbaert, 2019, 190 studies, 18,573 readers · Artificial Analysis, read on 9 Oct 2026

04

Bandwidth

you to them, typing

4 bits/s

a human to you, speaking

39 bits/s

an AI to you

360 bits/s

you, reading

20 bits/s

× 9

Speech carries 39 bits a second in all 17 languages measured, and a human brain handles about 10. The other lines apply the same method (5 characters a word, 1 bit a character) to 52 words a minute, 238 words a minute and 96 tokens a second.

Sources : Coupé et al., Science Advances, 2019, 17 languages · Zheng & Meister, Neuron, 2025, the method and the 10 bits/s

05

Paying for the answer

human, 2 min at minimum wage

$0.24

AI, one roast on this site

$0.0001

× 2,400

The U.S. federal minimum wage is $7.25 an hour. A roast reads about 385 tokens and writes 70, at $0.15 and $0.60 per million. Two minutes is an assumption: the time to read your question and type three sentences.

Sources : U.S. Department of Labor, in effect since 24 July 2009 · DeepSeek, off-peak price, read on 9 Oct 2026

06

Interrupting someone

interrupted human

23 min 15 s

AI

0 s

23 min 15 s

The average time an interrupted person takes to get back to what they were doing, measured by following employees at their desks.

Sources : Gloria Mark, Gallup, 2006, University of California, Irvine

07

Being right

AI as good as the pro, or better

84.9%

the pro does better

15.1%

84.9%

On real tasks from 44 occupations, judged blind against a professional's work, measured at the release of GPT-5.5. For the wifi password and what they are doing tonight, the human stays unbeaten.

Sources : OpenAI GDPval, via Decrypt, 23 April 2026

08

Everyone else already does

United States, adults

49%

France, aged 12 and over

48%

1 in 2

The share of people who use AI chatbots. In the United States, one adult in four uses one every day.

Sources : Pew Research Center, 5,119 adults, February 2026 · Arcep, Crédoc, 4,145 people, June 2025