Did you know? At 5 o'clock in the morning, the most common question people ask AI is how to fall asleep. At 7 o'clock in the morning, something big happened in the world. At 6pm, it’s time to decide what to eat tonight. Just last night, Anthropic released the sixth report in the economic index series - for the first time, the sampling accuracy of millions of Claude conversations was raised from weekly to hourly!
How anxious you are, how greedy you are, and what time you can't sleep are all in the data.
AI understands your daily schedule better than your partner.
When you are anxious or greedy, AI knows better than your partner.
First, there are weekdays and weekends.
Monday to Friday: business emails, PPT, marketing copywriting.
Saturday/Sunday: Emotional support, medical issues, investment advice.
In Claude's conversations, the proportion of personal use remains stable at around 35% during working days. By the weekend, it jumps directly to nearly 50%.
The usage of Claude Code is also changing. Back-end architecture, API debugging, and data storage all declined over the weekend, replaced by AI Agent design, quantitative trading, and game development.
The same group of people work for five days and be themselves for two days.
However, this "being oneself" does not always mean lying down.
Weekend startup-related conversations were at their highest level in all countries during the week, but job hunting activities fell along with other work tasks.
Weekends are a day to dream of being a boss, not a day to submit resumes.
Then, there are 24 hours a day.
Anthropic drew a frequency chart of different types of conversations by hour, which can be called an electrocardiogram of the rhythm of human life——
7am, news. 10-11 a.m. is the peak time for writing emails. At 6 p.m., I checked the menu and saw the largest increase in a single category throughout the day. In the evening, recommendations for watching dramas poured in. Around 5 a.m., the insomniac came.
Gardening talk, by contrast, barely moves from sunrise to sunset.
Anthropic couldn’t help but bury a joke in the report, calling gardening a “perennial topic of interest”—both a “perennial hot topic” and a “perennial plant.”
There is another layer of information hidden in the data after get off work and weekends: the work tasks handled by Claude are obviously tilted towards high-paying occupations.
The proportion of after-work conversations for low-paying positions such as secretarial and telephone sales dropped, but the proportion of high-paying positions such as marketing managers and programmers increased.
Highly paid workers have no time off work. This is not a new conclusion, but now there is hour-by-hour data to back it up.
Of course, the most dramatic part is tax day.
April 14, tax-related conversations are the daily average for May8 times. It was still high on April 15. On April 16, it fell off a cliff.
The American people collectively flocked to AI to ask about tax returns the day before the deadline. After that point, it could run faster than anyone else.
I write PPT during the day and fanfiction at night
In this report, Anthropic also introduced a new analysis dimension: artifact.
The thing you took away after talking to Claude, a document, a piece of code, an explanation, an email, all count.
93% of conversations produce artifacts. Only 7% are pure chats, leaving nothing behind.
Among them, the top three are: explanations (17%), documents and reports (15%), and guidance and suggestions (11%).
Overall, conversational output and written deliverables account for about one-third each, and coding and technical work account for one-sixth.
After classifying, Anthropic asked another question: Are these outputs at work or in life?
The answer varies by category.
81% of blogs and articles are for work.
Creative writing is just the opposite, with more than 80% being for personal use, with the main focus being fan fiction, worldview construction, and poetry; the remaining 13% of work scenarios are short video scripts and speech scripts.
Translation is the most "neutral", with 42% being work and 44% being personal. Planning is similar, with 44% working (entrepreneurial strategy, content strategy) and 49% personal (travel itinerary, fitness plan).
During the day, it’s the productivity engine. At night, I am a life assistant.
The higher the wage, the harder the AI will work
What is more interesting is the relationship between token consumption and wages.
Anthropic matches each conversation to the most corresponding occupation, and then compares the median salary of that occupation.
As a result, a rule emerged: conversations in high-paying professions consume more tokens.
The amount of conversation tokens corresponding to a marketing manager (with an hourly salary of $80) is approximately 2.5 times that of an editor (with an hourly salary of $37).
For conversations about building a website, the amount of tokens exceeds 3 times the median. One paragraph of explanation uses only one-fifth of the median.
Moreover, high-paying users are not simply “thrown to AI”.
They output more per round (1.34 times), interact more frequently (1.53 times), and engage in deep thinking more frequently (34% vs. 31%).
Claude worked hard and was not idle.
Of course, Claude not only did more, he also did "higher".
The reading level of its answers is generally higher than that of users' questions, and the average education level is about 1 year higher.
The biggest gaps are in images and graphics (+2.6 years), games (+1.9 years), and websites/applications (+1.7 years).
But when it comes to writing for an audience, the gap almost disappears - blogging -0.1 years, academic papers +0.0 years, and emails +0.3 years.
The reason is that the prompt of this type of task itself carries the same level of text material as the final output. You ask it to write the email for you, and you write a draft yourself first, so your reading level is of course about the same.
A diary you didn’t intend to write
The so-called "rhythm" means that you just open the dialog box every day, ask a few questions, and take what you need.
But when those conversations are cut down to the hour, output is divided into more than 30 categories, and each interaction is matched to an occupation and salary range, the fragments fit together into an outline.
Insomnia at 5 a.m., anxiety about dinner at 6 p.m., the idea of starting a business that suddenly appeared on the weekend, and the low mood that came late at night.
Taken individually, these are just a few hundred irrelevant questions. But put together, it is a person's daily routine, emotional cycle and days.
You may not have fully told these things to the people around you. But you're all handed a dialog box.
93% of conversations result in something. On the other hand, 93% of the conversations also leave you alone.
Anthropic said this report is to see clearly how AI can enter economic life. But once the data is accurate down to the hour, it will reveal more than just the economy.
During the day, Claude is your partner. At 5 a.m., only it knows that you are still awake.