Anthropic Economic Index report: Cadences

Source: Anthropic · 2026-06-26

One year ago, most Claude usage took the form of a conversation between a user and an assistant. With the rapid growth of Claude Code and Cowork, Claude sessions now increasingly consist of long-running agentic tasks. Chat transcripts no longer fully capture how people are using AI, and our methods for studying Claude’s economic impacts have had to adapt.

Our new privacy-preserving telemetry, which continuously samples a slice of conversations every day, allows us to study daily and hourly patterns in usage, in contrast to the seven-day samples each previous Economic Index report drew on.

We find that Claude usage mirrors the workweek, with personal prompts spiking on the weekend. The hourly data captures within-day patterns—people most often ask for sleep advice around 5 a.m. and for recipes around 6 p.m. We also see usage reflecting key dates. For instance, tax-related requests surged just before the US filing deadline on April 15.

The share of chat and Cowork3 conversations categorized as personal use spikes from around 35% on weekdays to just under 50% on weekends during the sample period (Figure 1.1). Outside the workweek, users’ conversations shift from business correspondence, marketing copy, and slide decks to emotional support, medical questions, and investment advice. This shift is biggest for high-income countries.

Request clusters5 allow us to go one level deeper and see which specific Claude Code tasks swing most between weekdays and weekends. On weekends, the Claude Code usage clusters that fall the most include backend architecture, API debugging, and data storage. Those that increase the most include AI agent design, quant trading, and gaming.

Weekends may also create space for people to pursue new ventures. Across countries, conversations related to starting a business are highest on Saturday and Sunday.

People ask for news at 7 a.m. local time. Business correspondence (e.g., email drafting) traces the arc of the workday, with a slight peak at 10–11 a.m. One of the biggest spikes is recipe requests, which are 2.3 times more frequent at 6 p.m. compared to the average. Media recommendations are most concentrated in the evening, while people seek sleep advice in the few hours just before dawn.

The sample period for this report covers tax filing deadlines for people in the United States. Figure 1.4 shows a large spike in the share of tax-related conversations around the deadline. On April 14, tax-related clusters were eight times as common as on the average day in May and remained about as high on April 15. On April 16, they dropped sharply.

Our classifier identified 93% of Claude conversations as producing an artifact (Figure 2.1).9 The most common artifacts are explanations (17% of conversations), documents and reports (15%), and guidance (11%). Conversational outputs (like explanations or guidance) and written deliverables (like documents or presentations) each account for about a third of conversations; code and technical work (like apps or scripts) for about a sixth.

Our January Economic Index introduced a primitive that classifies each conversation as work, personal, or coursework. Here, we apply that split to the artifacts produced in Claude conversations (Figure 2.2).

Some categories of artifacts are almost always personal. More than 80% of conversations producing creative writing, guidance, and recipes were classified as personal. Within categories, the personal and work-related uses can look quite different. Personal creative writing, for instance, is dominated by fanfiction, worldbuilding, and poetry; the 13% that is work-related is mostly in the form of short-form video scripts, screenwriting, and speeches. Categories most likely to be work-related include creating marketing content (80%), creating blogs or articles (81%), and writing database queries (82%).

We can also flip the question. Instead of asking what each output is used for, we can ask what sort of artifacts work, personal, and coursework conversations each tend to produce. Work conversations most often produce documents and reports (20%), followed by explanations (9%), email drafts (7%), and analyses and summaries (6%). Coursework conversations look broadly similar, with documents and reports leading there too (21%), closely followed by explanations (20%), educational materials (11%), and academic papers (6%). In contrast—and unsurprisingly—only 6% of personal conversations produce a document. Instead, the most common results are explanations (25%) and recommendations (22%).

Producing these outputs requires compute, and we find that compute tends to scale with the value of the work. We measure each conversation's computational costs in tokens—the amount of text processed and generated, including Claude's internal reasoning—and compare across occupations by mapping each conversation's classified task to the occupation that typically performs it. Throughout this section, we restrict our analysis to work-related conversations.

The left panel of Figure 2.3 shows a positive relationship between the median conversation-level number of tokens and the median wage in mapped occupation.10 For example, marketing managers earn roughly twice as much as editors ($80 vs. $37 per hour) and conversations mapping to their tasks consume approximately 2.5 times as many tokens. Admittedly, the relationship is noisy, and there are notable outliers. Pharmacists, for example, earn nearly three times what statistical assistants do ($68 vs. $24 per hour), yet conversations mapped to pharmacist tasks use only about one twentieth as many tokens.

The tokens consumed to generate different types of artifacts tell a similar story. More complicated and valuable outputs tend to consume significantly more tokens than simpler outputs. For example, conversations about building apps use more than three times the tokens of the median conversation. On the other end of the spectrum, a typical explanation uses about a fifth of the tokens of the median conversation. About 44% of the wage gradient in token consumption is explained by output mix—higher wage occupations are more likely to produce compute-intensive artifacts.

We measure this on a 1-5 scale, from "none" to "extreme.” Tasks that are easy to describe or specify involve little autonomy: the lowest-autonomy outputs are math or calculations, translations, and Q&As. High-autonomy tasks are those that require selection among many possible choices, e.g., creating apps and websites, games, or presentations. Such work, which requires sustained judgment, has historically been difficult to automate. By comparing the level of autonomy in Claude chat and Cowork to Claude Code, we show that this is starting to change.

Across almost all types of outputs (26 of 31 outputs shown) the level of AI autonomy is higher on Claude Code than chat or Cowork.11 For example, conversations producing scripts and code snippets involve 0.53 points more autonomy (on average, on the 1-5 scale) when created with Claude Code than conversations producing the same output on chat or Cowork.

In general, artifact types with higher-reading-level outputs also have higher-reading-level prompts (a correlation of 0.87 across conversations). However, we also observe that in almost every category, Claude’s output is at a higher comprehension level than the prompt, by roughly one year of education on average. The gap is widest where users describe something to be built, such as image and graphics (+2.6 years), games (+1.9), and apps and websites (+1.7). Some of the gap may simply be register; prompts are often terse and informal, while Claude tends to reply in polished prose. However, the gap is near zero for audience-facing writing (blogs −0.1, academic papers +0.0, email +0.3), possibly because prompts typically draft language or source material written in the same register as the intended output.

We find that most respondents expect significant AI progress over the next year. While people’s perception of AI capabilities depends on their experience, where they live, and how exposed their job is to AI, their expectations about the pace of future progress are strikingly uniform, consistent with a “rising tide,” in which AI capabilities improve broadly.

Views on what that progress means for their own careers are less uniform. Early-career workers report that AI can do the highest share of their work and express the most concern about job loss. Yet—contrary to a common concern—the people who delegate to Claude the most are the most optimistic about their future labor market outcomes, and feel their skills are growing in value. And despite (or perhaps because of) their proximity to AI's frontier, the average respondent’s hopes for the next decade center not on replacement but on collaboration. They hope AI can preserve meaningful work and automate the drudgery, and that its gains will be shared widely.