Anthropic published its third Economic Index report on June 26. It is the most detailed look yet at how people actually use AI in their daily lives.
The report, titled Cadences, introduces hourly usage sampling, conversation-level output classification, and survey data from roughly 9,700 Claude users whose responses were linked to their actual behavior through a privacy-preserving system.
The headline: AI usage mirrors the workweek, people hand over way more autonomy in Claude Code than in chat, and contrary to a lot of hand-wringing, the people who delegate the most to AI are the most optimistic about their careers.
Here is what the data actually shows.
The Rhythms of Real Usage
Anthropic upgraded its data pipeline to sample continuously at the hourly level instead of pulling seven-day snapshots. That shift alone reveals how deeply Claude usage tracks the rhythms of daily life.
Personal use of Claude spikes from around 35% on weekdays to just under 50% on weekends. During the workweek, conversations skew toward business correspondence, marketing copy, and slide decks. On weekends, the mix shifts to emotional support, medical questions, and investment advice.
The weekend personal-use bump is biggest in high-income countries.
The hourly data gets even more specific. People ask for news at 7 a.m. local time. Business correspondence peaks at 10-11 a.m. Recipe requests spike at 6 p.m. — 2.3 times the average rate.
Media recommendations cluster in the evening. Sleep advice peaks in the few hours just before dawn, around 3 a.m.
Tax season shows up in the data too. Tax-related conversations were eight times more common on April 14 than the average day in May, and they stayed elevated through the April 15 filing deadline. By April 16, they dropped sharply.
Higher-wage occupations show up more on nights and weekends. Tasks that map to roles like marketing managers and software developers increase their share of total conversations outside traditional work hours, while lower-wage tasks decline.
That gap holds even when you remove computer and math jobs from the analysis.
What People Actually Take Away
Anthropic also introduced a new classifier that labels the output of each conversation. They found that 93% of Claude conversations produce a recognizable artifact.
The most common outputs are explanations (17% of conversations), documents and reports (15%), and guidance (11%). Conversational outputs like explanations and guidance account for about a third of all conversations.
Written deliverables like documents and presentations account for another third. Code and technical work make up about a sixth.
What counts as work versus personal varies a lot by output type. More than 80% of conversations producing creative writing, guidance, and recipes are personal.
Work conversations are dominated by marketing content (80% work-related), blogs or articles (81%), and database queries (82%). Some outputs split almost evenly — translation is 42% work and 44% personal.
The token cost of each conversation scales with the value of the work. Conversations that map to higher-wage occupations consume more tokens. Marketing managers earn roughly twice as much as editors ($80 vs $37 per hour), and their Claude conversations consume about 2.5 times as many tokens.
Building an app uses three times the tokens of the median conversation. A typical explanation uses about a fifth.
Here is the piece that matters for the automation debate: in conversations tied to higher-wage jobs, Claude produces more output per turn (1.34 times), users engage more (1.53 times as many turns), and extended thinking is used more frequently (34% of conversations versus 31%).
These move together — more production from Claude does not mean less from the user. That pattern looks more labor-augmenting than labor-displacing.
Autonomy Depends on the Tool
The biggest gap in how much control people give Claude depends on which product they use. Claude Code and Cowork sessions show an average of 0.37 points more autonomy on a 1-5 scale than chat conversations.
About two thirds of that gap comes from the same tasks being executed with more delegation on Claude Code. The median blog-producing chat session involves 13 rounds of back-and-forth; the median Claude Code session producing the same output contains a single human prompt.
That gap is not just about model choice. Even when comparing conversations served by the same model, Claude Code still shows 0.26 points more autonomy. The product itself, not the underlying model, drives how much control people hand over.
Anthropic also looked at reading level. Claude consistently responds at a roughly one-year-higher comprehension level than the prompt.
The gap is widest for image and graphics prompts (+2.6 years), games (+1.9), and apps and websites (+1.7). Some of that is just register — prompts are terse, Claude writes in polished prose. But the gap nearly disappears for audience-facing writing like blogs (-0.1 years) and emails (+0.3), suggesting that when people are already writing in the register they want back, Claude stays closer to their level.
The Survey: Heavy Delegators Are Optimistic
The third chapter is the most interesting. Anthropic linked survey responses from about 9,700 users to their actual Claude usage data through the Anthropic Economic Index Survey.
This is the first time the Economic Index has been able to ask people what they think and cross-reference it with what they actually do.
The headline finding: people who delegate the most to Claude are the most optimistic about their careers. Across all six dimensions measured — pay, job security, ability to find a new job, meaning, autonomy, and human interaction — users with higher automation shares reported more positive expected impacts from AI on their work.
This is not just selection bias from early adopters. The relationship holds when controlling for how long someone has been using Claude. The more someone hands entire tasks to AI, the more they report that AI has made their skills more valuable.
And despite the concern that offloading thinking causes skill atrophy, heavy delegators report learning at the same rate as everyone else.
Almost 6 in 10 respondents expect AI to be able to do a higher share of their tasks in 12 months than it can today. Over a third expect AI to be able to do most or nearly all of their work tasks next year.
But there are real anxieties. Respondents are more worried about job loss for others than for themselves, and most worried about their junior colleagues. Over a third rated the probability of a junior colleague losing their job in the next year as over 60%.
Early-career workers report the highest share of tasks AI can do and express the most concern.
The survey also reveals a striking gender gap. Women make up only 12% of the linked survey sample. Even after accounting for occupational differences, their automation share is 0.33 standard deviations lower than men’s, and their Claude Code share is 0.24 standard deviations lower.
Instead, women log more active time on chat, suggesting more iterative, collaborative engagement.
The most common hopes for the next decade are not about replacement. People want collaboration with AI on meaningful work, automation of the tedious parts, and shared prosperity from AI’s economic gains.
What This Actually Tells Us
This is the most data-rich picture yet of how AI is embedding into economic life. The hourly sampling catches real behavioral patterns that weekly snapshots miss. The artifact classifier makes Claude’s output legible in a way that raw usage counts cannot.
The survey-to-usage linkage lets us see that attitudes track behavior in ways that the public debate often gets wrong.
The biggest takeaway for me: the people who use AI the most intensively are not the ones most worried about it. That does not mean the worries are unfounded — early-career workers and lower-wage occupations have real exposure. But the data does not support the narrative that heavy AI adoption is a recipe for deskilling or displacement.
The full report and PDF appendix with detailed methodology are available on Anthropic’s research page. If you use Claude daily, the patterns in this report are probably familiar. The numbers just put a scale on them.
