Anthropic switched from seven-day snapshots to continuous daily sampling, labeled the artifact of every conversation across 30+ categories, and added a survey of ~9,700 people. The two most counterintuitive findings have nothing to do with "how much time AI saves."
A year ago, most Claude usage was a back-and-forth between a user and an assistant. With Claude Code and Cowork, sessions are increasingly long-running agentic tasks, and chat transcripts no longer capture what people are actually using AI for.
This edition makes three changes to sharpen the picture: it samples at a higher rate, down to the hourly level (previous editions drew on seven-day snapshots); it adds a classifier that labels the artifact of each conversation; and it breaks out chat and Cowork conversations from the first-party API, aggregated monthly.
There is also a larger addition. Anthropic had lacked a view of Claude's impact outside of sessions: how people perceive AI changing their work and the opportunities in front of them, whether their usage shapes their expectations, and what they would want from AI in an ideal world. This edition asks directly, through a survey, and links the answers back to real usage data with privacy-preserving methods.
The report runs in three chapters: Cadences (when people use it), Artifacts (what they make with it), and Perceptions (how they feel about its effect on their work). The piece follows those three lines.
Continuous sampling makes a plain finding visible: Claude usage is close to a mirror of how people live. What people ask for at a given hour, and what they talk about on weekdays versus weekends, are etched into the logs.
The clearest signal is the weekday/weekend shift. Over the sample period, the share of chat and Cowork conversations classified as personal use rises from around 35% on weekdays to just under 50% on weekends. Outside the workweek, conversations move from business correspondence, marketing copy, and slide decks to emotional support, medical questions, and investment advice. The shift is biggest for high-income countries.
Going one level deeper, request clusters show which Claude Code tasks swing most. On weekends, the clusters that fall the most are backend architecture, API debugging, and data storage; those that rise the most are AI agent design, quant trading, and gaming. Conversations about starting a business also peak on weekends, while job-application activity drops along with other work tasks.
There is a within-day rhythm too. People ask for news at 7 a.m. local time; business correspondence traces the arc of the workday with a slight peak at 10–11 a.m.; one of the biggest spikes is recipe requests, 2.3 times more frequent at 6 p.m. than the daily average; media recommendations cluster in the evening, and sleep advice in the hours just before dawn.
At night and on weekends, when people do turn to Claude for work, the tasks skew toward higher-wage occupations — marketing managers and programmers are more likely to work off-hours, while the bottom two wage quartiles (telemarketing, clerical work) fall as a share. Removing computer and mathematical occupations in a robustness check leaves the pattern intact.
The sample period covers the US filing deadline. On April 14, tax-related conversations were eight times as common as on the average day in May, stayed high on April 15, and dropped sharply on April 16. A deadline on the calendar, printed exactly into the usage curve.
This edition classifies each chat and Cowork conversation by its artifact — a document, an explanation, a piece of code, a paper. 93% of conversations produce an identifiable artifact, most often explanations, documents and reports, and guidance.
Conversational outputs (explanations, guidance) and written deliverables (documents, presentations) each account for about a third of conversations; code and technical work for about a sixth. The same artifact can serve very different purposes: more than 80% of creative writing, guidance, and recipes are personal, while creating marketing content, blogs, and database queries are over 80% work. Plenty sit in the middle — plans and translation split roughly evenly between work and personal.
Compute tends to scale with the value of the work. Mapping each conversation to the occupation that typically performs the task: marketing managers earn roughly twice what editors do ($80 vs $37 per hour), and their conversations consume about 2.5 times as many tokens. App-building conversations use more than three times the median; a typical explanation uses about a fifth. About 44% of the wage gradient in token use is explained by the mix of output types.
In conversations mapped to higher-wage occupations, Claude produces more (1.34× output per turn), users engage more (1.53× turns), and extended thinking is enabled more often (34% vs 31%). These move together: more from Claude does not mean less from the user. With the human in the highest-value tasks, the pattern looks labor-augmenting rather than labor-displacing.
The median chat conversation producing a blog post involves 13 rounds of back-and-forth, while the median Claude Code session contains a single human prompt.Economic Index report: Cadences · Chapter 2
Autonomy is rated 1 to 5, from "none" to "extreme." Comparing chat/Cowork to Claude Code surfaces a counterintuitive result: for the same task, switching products changes how much the AI is allowed to decide — even when the underlying model is the same.
Across 26 of the 31 output types shown, autonomy is higher on Claude Code than on chat or Cowork. The sharpest example is the blog post: the requests behind it are similar on both surfaces, but the way people work differs entirely.
One might attribute this to model choice — Claude Code does run on the most capable models far more often (54% served by Opus, vs 10% on chat and Cowork). But the gap holds when the comparison is restricted to one model: among Sonnet conversations, Claude Code still shows 0.26 points more autonomy. Across all conversations the gap is 0.37 points, about two-thirds of which comes from the same tasks being executed with more delegation. The product used is likely more important than the underlying model.
For each conversation, a classifier estimates two reading levels — the prompt and the response — in years of education needed to understand the text. In almost every category, Claude's response sits about one school year above the prompt. The gap is widest where users describe something to be built.
For audience-facing writing the gap is near zero (blogs −0.1, academic papers +0.0, email +0.3), likely because those prompts are already drafted in the register of the output. The wide gaps are all "describe it, have the AI build it" work — the person gives a line, Claude supplies the rest.
The first two chapters cover how people use Claude; this one covers how they feel. The survey links ~9,700 respondents to real usage data. The sample is not representative — computer and mathematical occupations are ~30% of respondents (4% of US employment), management is 23% (7% of employment), and physical occupations are under-represented.
Most respondents expect significant progress over the next year: close to 6 in 10 picked a higher band for next year than for today, and over a third expect AI to handle most or nearly all of their work tasks. And regardless of how exposed their occupation is, expectations of the pace of progress are strikingly uniform — a software engineer and a construction manager anticipate roughly the same increment within their field.
The counterintuitive part is below. Researchers split expected one-year impact into six dimensions — pay, job security, ability to find a new job, plus meaning, autonomy, and human interaction — and tracked how each varies with automation share (how much someone hands entire tasks to Claude). The result: across all six dimensions, the more automated users are more optimistic, with the largest effects on expected pay and ability to find a job.
Who is anxious? Early-career workers: they report AI can do the highest share of their work and worry most about job loss. A common concern is that handing entire tasks to AI offloads thinking and erodes skill. The data does not show that pattern — heavy delegators report learning at the same rate as everyone else, and are more likely to feel their skills are growing in value. But the researchers are explicit: these are self-assessments and do not rule out skill erosion.
Worry is real, but it lands on others rather than oneself. Over a third think their own responsibilities will change significantly; 10% think they will likely lose their own job (38% of those attribute the forecast to AI). Yet over a third put a junior colleague's odds of job loss next year above 60%. There is also a clear gender difference — women are only 12% of the linked sample, and even controlling for occupation, their Claude Code share is 6.3 points lower and automation share 7.3 points lower, with more iterative use.
The survey closes with an open question: what do you hope an AI-shaped economy looks like in ten years? The most common answers do not land on "replacement": over half want to collaborate with AI on meaningful work, just over half want AI to automate the tedious parts and return free time, and about a third hope the gains are widely shared.
anthropic.com · 2026-06-26 · by Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig and others. All claims and figures here come from that report. Charts are site components redrawn from the reported values: the hourly cadence chart and the six-dimension optimism chart use bar lengths only to show direction (the report gives no exact coefficients, as noted in the captions); all other shares, multiples, and point values are the report's own. The survey sample is not representative and is self-reported; the report itself notes the selection caveat and that skill erosion cannot be ruled out.