Blog › Tech · Field Notes
28 million words in 60 days
What heavy AI use actually looks like
A usage dashboard told us we had used 155 times more text than the novel Dune. We went looking for the research to work out whether that meant anything. It does, but not the thing we expected, and not the thing another AI told us.
The short version
Six lines, and then the working-out
- Our card said 37.9 million tokens across 30,449 messages in 60 active days. That is roughly 28 million words, or about 950 times the length of Animal Farm.
- Daily AI use is no longer rare. Around 24% of US adults use an AI chatbot every day, and 58% of Australians aged 14 and over used an AI tool in the last four weeks.
- What is rare is depth. Against the only published per-user figure we could find, the average is about 26 messages a week. We ran roughly 2,370.
- Most of those messages were not typed by a human. They were agentic turns - the model reading files, running tests and correcting itself. We say so before anyone else does.
- Volume proves range, not quality. A big number tells you someone showed up. It tells you nothing about whether the work was any good.
- We still hold the keys. No connected inbox, no connected bank, no autonomous agents, and a human approves every production write.
It started with a comparison we did not understand
A dashboard, a number, and a novel we had to go and look up.
Anthropic put a usage card in the Claude interface. Ours read: 30,449 messages, 37.9 million total tokens, 60 active days, longest streak 24 days, peak hour 7pm. Underneath it, a line comparing our text volume to a book: 155 times more tokens than Dune.
We had no idea what that meant. Not the number - the book. Dune turned out to be Frank Herbert's science fiction novel, roughly 188,000 words, which is about 250,000 tokens once you convert it into the chunks a language model actually reads. So the card was saying we had pushed a hundred and fifty-odd novels' worth of text through one tool in two months of active days.
That is a strange thing to read about yourself on a Tuesday night. It is also, on its own, meaningless. A number without a benchmark is a vanity metric, and we build things for a living precisely so we do not have to run on vanity metrics. So we went and found the research.
Anthropic's usage summary, August 2026.
Source: Anthropic Claude usage summary, screenshot August 2026
~28.4M
Words, once 37.9M tokens are converted at the standard 0.75 words per token
Our card, our arithmetic
1,245
Average tokens per message. This is the number that says documents, not chat
37.9M / 30,449
507
Messages per active day, across 60 active days
30,449 / 60
28 million words, measured in books
Because nobody has an instinct for what a million tokens feels like.
Animal Farm
~30,000 words
×949
Dune
~188,000 words
×151
Complete Shakespeare
~885,000 words
×32
Multiples calculated against ~28.4 million words. Novel word counts are approximate and vary by edition. Anthropic's own card independently reports 155x Dune measured on tokens.
George Orwell wrote Animal Farm as a novella you can read in an afternoon. We put roughly nine hundred and fifty of them through one tool. Herbert's Dune is the one people describe as dense and enormous. A hundred and fifty-one of those.
There is a real sense of wonder in that, and we are not going to pretend otherwise. But wonder is where the thinking starts, not where it stops. The next question is the only one that matters: compared to whom?
First, the part everybody gets wrong
Using AI every day stopped being remarkable some time last year.
When we asked another model where these numbers sat, it told us we were in the "top 1 to 2% of users globally" and offered a tidy table of user tiers to prove it. It was a satisfying answer. It is also, as far as we can establish, unsourceable. No AI provider publishes per-user token percentiles. Nobody outside those companies can calculate that number, which means nobody outside those companies should be quoting it, including us.
What is published is adoption data, and it says something more interesting than a percentile would. Daily AI use has quietly become normal.
Share of each population group, most recent published figures.
Sources: Roy Morgan, March 2026 (n=14,646) · OECD ICT Access and Usage Database via ITIF, 2025 figures · Pew Research Center, June 2026 (n=5,119) · Gallup, May 2026. Measures differ - read each bar against its own label, not against the others.
Roughly one US adult in four now opens an AI chatbot every day, from a Pew Research Center survey of 5,119 people. So "I use AI daily" is not a credential. It is closer to saying you check your email.
Australia sits well ahead of the OECD average on this. Roy Morgan found 13.6 million Australians aged 14 and over - 58% - used an AI tool in an average four weeks to March 2026, from a sample of 14,646. In the 25 to 34 bracket it is 74%. Across OECD countries the 2025 figure was 36.8% of individuals. We are, as a country, early adopters.
So if frequency is not the thing that separates users any more, what is? Two things: how much you put in, and how many different kinds of work you put in.
The only honest comparison we could find
One published per-user figure exists. Here is how it stacks up, caveats and all.
The clearest per-user number in public comes from OpenAI's own research team, published through the National Bureau of Economic Research: How People Use ChatGPT. By July 2025 the platform had around 700 million weekly active users sending about 18 billion messages a week.
Divide one by the other and the average user sends about 26 messages a week. Under four a day.
26
Messages per week, average user
NBER 34255
~2,370
Messages per week, ours, across the same period
Our card, our arithmetic
90-140×
The multiple, depending on whether you count calendar weeks or active days
Both methods shown below
We are giving you a range rather than a single clean number, because the two defensible ways of calculating it disagree. Per active day, our 507 against the average 3.7 is about 138 times. Spread across calendar weeks, our ~2,370 against 26 is about 92 times. Both are true. Neither is the "real" one.
And here is the caveat that matters most: that baseline is ChatGPT and our figures are Claude. It is the best public per-user comparison that exists, and it is still not like for like. We would rather say that out loud than have you find it yourself and wonder what else we skipped.
Most of those 30,449 messages were not typed by a human
The part a technical reader would go for, so we are going there first.
If you have used an agentic coding tool, you already know what is coming. Those tools run their own loop. You give one instruction, and the model then reads files, writes code, runs the tests, reads the failure, fixes it, runs them again. Every one of those steps is counted as a message.
Anthropic measured exactly this. In its June 2026 Economic Index report, the median chat conversation that produced a blog post involved 13 rounds of back and forth. The median coding session that produced the same thing contained a single human prompt.
So what is the message count actually measuring
Not typing. It is measuring how many autonomous cycles were set running and supervised. That is a different skill from typing fast, and it is the one worth having - but it is not the skill the raw number implies, and we would be misleading you to let it stand.
Which is why the token figure is the more honest signal. 1,245 tokens per message, on average. A chat message is a sentence or two. Twelve hundred tokens is a file. Across 30,449 messages, that is codebases, database schemas, error logs, survey exports and full documents going in and coming back changed.
That is the difference between using AI as a search box and using it as somewhere the actual work lives.
The lanes, and why there is more than one
Most heavy users specialise. The volume above comes from refusing to.
Anthropic's data shows how concentrated AI use is: computer and mathematical tasks account for about a third of consumer traffic and nearly half of API traffic. Most heavy users are heavy in one direction.
Our 37.9 million tokens are spread across work that has almost nothing in common except that it all had to ship. Every item below links to the post where we wrote it up, mistakes included:
Production web build
This site. Astro, Tailwind v4, GSAP scroll animation, art-directed video that swaps source by viewport, a static build on a CDN. Not a template with the colours changed.
Micro-SaaS, end to end
Scoping, build, payments, launch. Including the bit where the payment flow broke on launch day and we had to debug it live.
Document engineering
Rebranding 550 documents in eleven minutes meant learning how Word files are actually built - headers, footers, tables, the XML underneath. Then testing it properly before letting it near anything real.
Data and analytics
Retention modelling, survey sentiment analysis, financial calculators with the arithmetic checked by hand against published rates.
Knowing when not to use AI at all
The one we would point at first. We wrote a whole piece on choosing deterministic Python over an AI wrapper, because for a task that must give the same answer every time, a language model is the wrong tool and no amount of enthusiasm changes that.
That last one is the point of this whole section. There is a version of heavy AI use that is really just clicking accept on things you cannot read. It produces impressive demos and unmaintainable software, and it falls over the first time something breaks in production at 11pm.
The distinction is not how much code you generate. It is whether you can read it, test it, and tell when the confident answer is the wrong one.
What we have not connected, and why
The volume is only trustworthy if the controls are.
Here is the part we think matters more than any of the numbers above. Thirty-eight million tokens of AI use tells you nothing about whether the person running it is careful. So here is the boundary, written down, in the same terms we would give a client.
Connected inbox
Our email is not wired to any AI tool. An inbox is the master key to every other account you own. The convenience has never been worth it.
Connected banking or financial accounts
Nothing that moves money is connected to anything that generates text. That line is not up for review.
Autonomous agents running unattended
We use agentic tools, heavily. We do not leave them running against live systems with nobody watching. Delegation is not the same as absence.
A human approves every production write
Reads are free. Anything that changes a live database, a live site or a customer record is reviewed and approved by a person first. Every time, including when it is slow and annoying.
The person driving reads the output
Not every line. Enough to know what changed and why it changed. The moment you stop reading is the moment you cannot tell good work from confident work.
None of this is a boast about caution. It is the reason the volume is worth anything. Anyone can generate 38 million tokens. The question a client should ask is what that generation was allowed to touch, and who checked it before it went live. If a supplier cannot answer that in one page, the volume is a liability rather than a credential.
If you want the longer version of how we think about this, we wrote it up in is Claude safe for work.
What the number proves, and what it does not
The unglamorous close, which is the only kind worth writing.
Strip away everything that does not survive a source check and you are left with something smaller and more useful than the headline. Someone put an enormous amount of real material through these tools, across an unusual number of different problems, over a sustained period, and shipped things at the end of it that are running in production now.
That is a claim about range and persistence. It is not a claim about talent, and volume has never been a proxy for quality. Anyone can burn tokens. The work either runs or it does not.
One line worth adding, and then we will leave it. These are a woman's numbers. A Harvard Business School review of 76 studies covering more than 300,000 people found the depth-of-use gap between women and men has been stuck near 16% since early 2025, and Pew puts daily chatbot use at 27% of men against 20% of women. We were not trying to make a point about that. There was just work to do.
Generative AI adoption, and use of internal AI coding tools within one employer.
Source: Cranney, Delecourt and Koning, Harvard Business School Working Paper 25-023, May 2026. Coding-tool figures are software engineers at a single employer.
Five things to do with this
All doable this week, none of them requiring 38 million tokens.
Go and find your own numbers
Most AI tools now publish a usage view. Look at tokens as well as messages. The ratio between them tells you more about how you work than either number alone.
Count tokens, not conversations
A high token-per-message figure means you are feeding real material in - documents, data, code. A low one means you are chatting. Neither is wrong. They are just different jobs.
Write down what you will never connect
Decide it once, in writing, while nothing is on fire. Ours is short: no inbox, no banking, no production write without a human approving it.
Add a second lane on purpose
If everything you do with AI is one task type, you are only seeing a slice of what it does. Pick something adjacent to your day job and build one small thing end to end.
Read the code it gives you
Not all of it, not line by line. Enough to know what it changed and why. The moment you stop reading is the moment you stop being able to tell good output from confident output.
See what the volume was actually for
The tools, kits and calculators built during those 60 days. Free to use, no sign-up.
Open the free toolsQuestions people ask
Short answers, each one standing on its own.
What counts as an AI power user?
There is no official definition, and no AI provider publishes per-user token percentiles, so anyone claiming you are in the 'top 1%' is guessing. A more useful measure is intensity and breadth: how much material you feed in per message, and how many different kinds of work you use it for.
How many messages does the average AI user send?
The clearest public figure comes from OpenAI's own research. By July 2025 ChatGPT had around 700 million weekly active users sending about 18 billion messages a week. That works out to roughly 26 messages per user per week, or under four a day.
Does high token usage mean better results?
No. Tokens measure volume, not value. High token use usually means you are feeding in real documents, data or code rather than chatting, which is a signal about the kind of work being done. It is not a signal about whether the work is correct.
Should I connect AI tools to my email and bank accounts?
That is a risk decision, not a technical one. We have chosen not to. Connecting an inbox or a financial account hands over standing access to data you cannot easily un-share, and the convenience rarely justifies it for a small business. Start with read-only access to things you can afford to expose.
Why do agentic AI tools produce such high message counts?
Agentic tools run their own loop - reading files, calling tools, running tests and correcting themselves - and each of those steps is counted as a message. Anthropic's own research found the median coding session that produced a blog post contained a single human prompt. So a large message count reflects delegated cycles, not typing.
How many Australians use AI?
Roy Morgan found that 13.6 million Australians aged 14 and over, or 58%, used an AI tool in an average four weeks to March 2026. Uptake peaks in the 25 to 49 bracket at over 70%. That puts Australia ahead of the OECD average of 36.8% for 2025.
Is there a gender gap in how people use AI?
Yes, and it is about depth as well as adoption. A Harvard Business School review of 76 studies covering more than 300,000 people found men adopt at 47.8% against 39.3% for women, a relative gap that has sat near 16% since early 2025. Pew Research Center found the same pattern in daily use: 27% of US men against 20% of US women. Women's sessions also run shorter across the major tools.
Sources
- Roy Morgan, Artificial Intelligence (AI) tools usage, March 2026 (n=14,646)
- OECD, ICT Access and Usage Database, 2025 figures, via ITIF, March 2026
- Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman, How People Use ChatGPT, NBER Working Paper 34255
- Cranney, Delecourt and Koning, Global Evidence on Gender Gaps and Generative AI Over Time, Harvard Business School Working Paper 25-023, May 2026
- Anthropic Economic Index report: Cadences, June 2026
- Anthropic Economic Index report: Learning curves, March 2026
- Gallup, Gen Z's AI Adoption Steady, but Skepticism Climbs, February to March 2026
- Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact, 17 June 2026 (n=5,119 US adults, surveyed 17-23 February 2026)
- Pew Research Center, The gender gap in AI in the US, 17 June 2026
Written from our own usage card and eight published sources, every one linked above. All arithmetic recomputed by hand from the raw figures rather than taken from any model's summary - which is how the "top 1 to 2%" claim came out of the draft. Charts are inline SVG in the site's own palette, no chart library. Researched and drafted with Claude (Fable and Opus), edited by a person. Errors are ours.