Anthropic Economic Index Cadences: How People Really Use Claude AI

Introduction

Anthropic just dropped one of the most fascinating pieces of AI research this year, and it has nothing to do with model benchmarks or context windows. The June 2026 Anthropic Economic Index report, titled "Cadences," is a deep dive into how millions of people actually use Claude AI throughout their days, what they produce with it, and how they feel about AI reshaping their careers. The findings are surprising, sometimes counterintuitive, and essential reading for anyone who uses Claude regularly.

This report marks a significant evolution in how Anthropic studies its own product's economic impact. With Claude Code and Cowork transforming sessions into long-running agentic tasks rather than simple chat exchanges, the old methods of analyzing transcripts no longer capture the full picture. Anthropic responded by overhauling its data pipeline, sampling at hourly rates, introducing new output classifiers, and launching a survey of nearly 10,000 linked respondents.

Let's break down what they found and what it means for Claude power users.

The Rhythms of Claude Usage Mirror Everyday Life

The headline finding from the "Cadences" chapter is elegantly simple: Claude usage follows the rhythms of real life with remarkable precision.

Anthropic's new hourly telemetry data reveals that work-related queries dominate weekdays, with personal use sitting at about 35% of conversations. Come Saturday and Sunday, personal use spikes to nearly 50%. The types of conversations shift dramatically too. During the workweek, people draft business correspondence, create marketing copy, and build slide decks. On weekends, they pivot to emotional support, medical questions, and investment advice.

The hourly data is where things get particularly interesting. People ask Claude for news at 7 a.m. Business email drafting traces the arc of a typical workday, peaking around 10 to 11 a.m. Recipe requests spike at 2.3 times their average frequency at 6 p.m., right when people start thinking about dinner. Media recommendations cluster in the evening hours. And in perhaps the most human finding of the report, sleep advice requests peak around 5 a.m., when insomnia drives people to their screens.

There are also calendar-driven surges. Tax-related Claude conversations spiked eight times above their average on April 14, the day before the US filing deadline. By April 16, they had dropped back to baseline. Claude has become so embedded in people's routines that the tax deadline shows up as clearly in usage data as it does on IRS servers.

Who Works Off-Hours With Claude

One of the more revealing findings concerns who uses Claude outside traditional work hours. When people turn to Claude for work on nights and weekends, the tasks skew heavily toward higher-wage occupations. Marketing managers, computer programmers, and other high-earning professionals are the most likely to be doing work-related Claude tasks outside the nine-to-five.

Meanwhile, tasks related to lower-wage occupations like telemarketing and clerical work shrink as a share of total conversations during off-hours. This pattern held even when Anthropic removed all computer and mathematical occupations from the analysis, confirming it is not simply a reflection of developers coding at night.

Weekends also appear to create space for entrepreneurial ambition. Conversations related to starting a business peak on Saturdays and Sundays across countries. People are using their free time not just for personal queries but to explore new ventures, brainstorm business models, and plan side projects with Claude as their thinking partner.

What People Actually Produce With Claude

The report introduces a new concept called "artifacts," referring to the primary output Claude produces in a conversation, whether that is a document, explanation, piece of code, or academic paper. A classifier identified 93% of Claude conversations as producing some kind of artifact.

The most common outputs break down as follows. Explanations lead at 17% of conversations, followed by documents and reports at 15%, and guidance at 11%. Conversational outputs like explanations and recommendations account for about a third of all conversations. Written deliverables like documents and presentations make up another third. Code and technical work cover roughly a sixth.

The split between work and personal use varies dramatically by artifact type. Over 80% of creative writing, guidance, and recipe conversations are personal. Personal creative writing is dominated by fanfiction, worldbuilding, and poetry, while the small portion that is work-related tends to be video scripts, screenwriting, and speeches. On the work side, blog and article creation is 81% work-related, marketing content hits 80%, and database queries reach 82%.

Some artifact types sit right in the middle. Plans and strategies split almost evenly between work at 44% and personal at 49%. Personal planning centers on travel itineraries and workout schedules, while work-related planning focuses on entrepreneurial and content strategies.

More Valuable Work Requires More Compute

Anthropic found a clear positive relationship between the economic value of work and the computational resources consumed to produce it. Conversations mapped to higher-wage occupations use significantly more tokens than those mapped to lower-wage jobs.

For example, marketing managers earn roughly twice what editors do, and their Claude conversations consume approximately 2.5 times as many tokens. The pattern makes intuitive sense: more complex, higher-value work requires more reasoning, more extended thinking, and more iterative refinement from Claude.

The artifact types that consume the most compute are also the most complex. Building apps uses more than three times the tokens of a median conversation, while a typical explanation uses about a fifth. About 44% of the wage gradient in token consumption is explained by output mix, meaning higher-wage occupations are simply more likely to produce compute-intensive artifacts.

Crucially, in higher-value conversations, both Claude and the human user contribute more. Claude produces 1.34 times as much output per turn, while users engage in 1.53 times as many turns and enable extended thinking more frequently. This pattern suggests augmentation rather than replacement: the human stays involved in high-value tasks, and the collaboration deepens rather than diminishes.

Claude Code Users Delegate Far More Autonomy

The report measures AI autonomy on a one-to-five scale, from tasks where Claude has no discretion to those requiring extreme judgment. The findings reveal a striking gap between surfaces.

Across almost all output types, 26 out of 31 categories, the level of AI autonomy is higher on Claude Code than on chat or Cowork. The average difference is 0.37 points on the five-point scale, which is substantial. For scripts and code snippets, the gap widens to 0.53 points.

About two-thirds of this difference comes from the same tasks being executed with more delegation on Claude Code. Blog posts illustrate this perfectly: the requests behind them are similar across surfaces, but the workflow differs sharply. The median chat conversation producing a blog post involves 13 rounds of back-and-forth. The median Claude Code session producing the same output contains a single human prompt.

Interestingly, this gap persists even when comparing conversations served by the same model. Among Sonnet-only conversations, Claude Code sessions still show 0.26 points more autonomy. The product matters more than the model, a finding that should interest anyone choosing between chat and agentic workflows.

Claude Responds Above the Level It Is Asked

Another fascinating finding: Claude consistently produces output at a higher reading level than the prompt it receives. On average, Claude's responses require about one additional year of education to comprehend compared to the user's input.

The gap is widest for creative and constructive tasks. Image and graphics prompts get responses 2.6 years above their input level. Games show a 1.9-year gap. Apps and websites sit at 1.7 years. Some of this is simply register, since prompts tend to be terse and informal while Claude replies in more polished prose.

But the gap nearly vanishes for audience-facing writing. Blog posts show a negative 0.1-year difference, academic papers are at zero, and emails sit at plus 0.3 years. This makes sense because when people write prompts for audience-facing content, they tend to include draft language or source material written in the same register as the intended output.

The Survey: What Workers Really Think About AI

Chapter three of the report presents the first results from the Anthropic Economic Index Survey, linked to actual Claude usage data through privacy-preserving methods. The final sample includes about 9,700 respondents, heavily skewed toward computer and mathematical occupations at 30% and management at 23%.

The headline numbers on productivity are striking. 86% of respondents report speed gains, 82% report scope gains, and 69% report quality improvements. Beyond productivity, 68% say they learn more with AI, and 57% feel AI has made their skills more valuable.

But the most interesting findings concern expectations about the future. Close to 60% of respondents chose a higher capability band for AI in 12 months compared to today. Over a third expect AI to be able to do most or nearly all of their work tasks next year. That is a remarkable level of anticipated progress from people who use these tools daily.

The Automation Paradox: Delegators Are the Most Optimistic

Here is the most counterintuitive finding in the entire report, and arguably the most important. People who delegate the most to Claude, using it in the most automated way, are the most optimistic about their own job prospects.

Across all six dimensions measured, including pay, job security, ability to find a new job, meaning, autonomy, and human interaction, people with a higher share of automated sessions feel more positive about AI's impact on their careers. The largest effects appear in expectations about future pay and employability.

This runs directly counter to the intuition that heavy delegation leads to anxiety about replacement. Instead, the people closest to AI's frontier seem to see the most opportunity in it. Anthropic suggests two possible explanations. Selection effects may play a role, meaning the most enthusiastic users are naturally the most willing to delegate. But the pattern holds even when controlling for user tenure on Claude, which serves as a proxy for early-adopter enthusiasm.

The alternative explanation is that automated usage provides direct experience of AI's benefits. If you hand Claude an entire task and see excellent results, you learn firsthand what AI can do for your productivity, and that experience breeds confidence rather than fear.

Skill perception follows the same pattern. The share of people reporting that AI increases the market value of their skills rises with automation share. Meanwhile, self-reported learning remains flat across all levels of delegation, contradicting the concern that heavy delegators stop learning.

Early-Career Workers Feel the Most Exposed

Not all the survey data is optimistic. Perceptions of AI exposure are negatively correlated with years of work experience. People with at least 15 years of experience estimate that AI can do about 10 percentage points less of their work compared to those in their first year.

When asked what tasks AI could never do, experienced workers emphasized judgment, contextual awareness, situational reasoning, and the relational dimensions of their jobs, things like building trust and managing people. These are the tacit skills that accumulate over a career and resist automation.

Respondents are also more worried about job loss for others than for themselves, a pattern psychologists call the "better-than-average effect." Over one-third of respondents rated the probability of a junior colleague losing their job in the next year at above 60%. Meanwhile, only 10% rated their own job loss as likely or very likely.

Gender Differences in Claude Usage

The survey reveals significant gender differences in how people use Claude, even after controlling for occupation. Women, who make up 12% of the linked respondent sample, are marginally less likely to use Claude for work. Their share of Claude Code sessions is 0.24 standard deviations lower, and their automation share is 0.33 standard deviations lower.

Instead, women tend to use Claude more iteratively and collaboratively, logging more active time on chat than men. This is not simply a function of occupational differences. The patterns persist after conditioning on occupation, suggesting genuine differences in interaction style that deserve further study.

What People Hope For From AI

The survey ends on a forward-looking note, asking respondents to dream big about what an AI-shaped economy should look like in ten years. The top themes are illuminating.

Over half of respondents expressed a desire for AI augmentation of work, wanting to collaborate with AI on meaningful tasks while preserving the importance of their careers. Simultaneously, just over half hoped for automation of tedious work, freeing up time for meaning both inside and outside of work. About one-third of respondents emphasized shared prosperity, hoping the economic gains from AI would be distributed widely rather than concentrated at the top.

These aspirations paint a remarkably coherent picture. People do not want to be replaced. They want to be amplified. They want the boring parts automated and the meaningful parts preserved, with the benefits shared broadly.

What This Means for Claude Power Users

The Cadences report contains several practical takeaways for anyone who uses Claude regularly. First, the data confirms that Claude Code enables significantly more delegation than chat interfaces, even for the same tasks. If you are still doing everything through chat, you may be leaving productivity gains on the table.

Second, the finding that higher-value work consumes more compute suggests that investing in extended thinking and longer sessions pays off for complex tasks. Do not be afraid to let Claude reason deeply on important work.

Third, the automation paradox should reassure anyone worried that delegating too much to AI will hurt their career. The data suggests the opposite: people who delegate the most feel the most confident about their skills and job prospects.

Finally, the usage cadence data is a reminder that Claude has become deeply woven into the fabric of daily life for millions of people. From 5 a.m. insomnia queries to 6 p.m. recipe requests to weekend entrepreneurial brainstorming, Claude is not just a work tool. It is a companion for the full cycle of human activity.

Conclusion

The Anthropic Economic Index Cadences report is the most detailed look yet at how AI is diffusing into economic life. Its findings challenge simple narratives about AI replacement, showing instead a complex picture where augmentation dominates, heavy users are the most optimistic, and the rhythms of daily life shape AI usage just as much as the technology shapes daily life.

For those tracking their own Claude usage patterns, tools like Gaugr offer real-time visibility into consumption across models and sessions, helping you understand your own cadences alongside Anthropic's aggregate data.