Anthropic Economic Index report: Cadences

The Anthropic Economic Index report of June 2026 moves the index from one-week samples to continuous hourly sampling and adds a classifier for the output of each conversation; it finds usage following the working week, the day and the US tax deadline, token use rising with the wage of the mapped occupation, more delegation in Claude Code than in chat for the same outputs, and, in the first results of the Economic Index Survey (about 9,700 linked respondents, not representative), that users whose sessions are more automated expect AI to do more of their work and are more optimistic about its effects on their jobs.

Maxim Massenkoff, Eva Lyubich, Szymon Sacher, et al. (Anthropic) · 2026-06-26 · original

As Claude's use moved toward long-running agentic sessions in Claude Code and Cowork, the Anthropic Economic Index, which had classified one-week samples of transcripts, changed its methods. The report published on 26 June 2026, subtitled "Cadences," is the first on the new pipeline. Its authors are Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller and Peter McCrory of Anthropic's Economics team. Its three chapters cover usage by hour, weekday and date; what Claude produces, its token cost and how much is delegated; and the first results of the Anthropic Economic Index Survey, launched in April 2026, linked to respondents' own usage. It follows the March 2026 report (Learning curves).

This note follows the 33-page PDF and the 15-page appendix (26 June 2026); the web page has the same text with small wording differences and no figure numbers. The Hugging Face data release of the same date (release_2026_06_26) is used where noted.

What changed in the pipeline

The introduction and appendix Chapter 1 list the changes:

  • Sampling. A fixed number of conversations is sampled every hour, continuously, instead of a seven-day sample per report. This allows daily and hourly analysis.
  • Surfaces. Chat and Cowork conversations from consumer Free, Pro and Max accounts on Claude.ai and the desktop app are analysed together as "Claude conversations"; first-party (1P) API traffic, which excludes Claude Code and third-party clouds, is reported separately; Claude Code appears in some comparisons (introduction, footnote 1; Chapter 1, footnotes 2 and 3).
  • A new artifact classifier labels the single most prominent concrete output of each conversation.
  • A new O*NET classifier. Claude first picks a detailed work activity, then a task within it (O*NET 30.2), scoring candidates 1 to 5 for confidence; one of the highest-scoring tasks is chosen at random. Options are shuffled because "Claude may have order preferences," and the classifier sees more of each transcript than before.
  • New request clusters at the top level, and monthly aggregates in the public release.

All classification is done by Claude ("transcripts are only read by another instance of Claude"), and cells with too few observations are dropped (Chapter 1, footnote 1). The report does not name the classifying model. The appendix illustrates the classifiers on WildChat, a public dataset of ChatGPT conversations, including a question about Rust libraries mapped to a sales task that it says "might have been better assigned" to a software task; it reports no check of either classifier against human coding.

Chapter 1: usage over the week, the day and the year

The working week

The share of chat and Cowork conversations classified as personal rises from around 35% on weekdays to just under 50% on weekends (Figure 1.1, which plots late April to late May 2026). Claude Code and the 1P API show the same weekly pattern at lower levels. Conversations about starting a business peak on Saturday and Sunday, while job-application work falls with other work tasks.

The day

Figure 1.2 plots each request cluster's hourly share against its own daily average, by local time inferred from the IP address (footnote 6). News requests peak at 7 a.m., business correspondence rises through the working day with a slight peak at 10 to 11 a.m., recipe requests are 2.3 times their average at 6 p.m., media recommendations peak in the evening and sleep advice in the hours before dawn. The introduction and the web text put the sleep-advice peak "around 5 a.m.", where the tallest bar in the figure's panel is (read from the figure); the opening of Chapter 1 in the PDF says "around 3 a.m."

When people use Claude for work at night and on weekends, the tasks shift toward better-paid occupations. Figure 1.3 prints the percentage change in each wage quartile's share of work conversations, nights and weekends against weekday working hours: −4% for the lowest quartile, −11% for the second, +8% for the third and +8% for the highest. The pattern held when computer and mathematical occupations were removed. The report notes that it "can't conclusively identify the jobs of the people making these requests" and offers out-of-hours work in high-paid occupations as a possible explanation.

Tax day

Tax-related clusters in the US were "eight times as common as on the average day in May" on 14 April, about as high on 15 April, the filing deadline, and fell sharply on 16 April (text before Figure 1.4). Figure 1.4, which plots the share relative to its own May average and prints no values, peaks at about 6.8 on 14 April and 6.6 on 15 April (read from the figure), below the text's "eight times"; its series for traffic outside the US shows no spike.

Chapter 2: artifacts, compute and delegation

Chapter 2 covers chat and Cowork conversations sampled from 10 April to 10 June 2026, with Claude Code sessions from the same period where surfaces are compared (footnote 1 of the chapter). The artifact classifier chooses one label from 32 options, including "other" and "none" (appendix Chapter 2); the report speaks of "more than 30 categories."

What Claude produces

The classifier finds an artifact in 93% of conversations (Figure 2.1). The most common are explanations (17%), documents and reports (15%) and guidance (11%). On the 1P API, 16% of conversations have no clear output, against 7% on chat and Cowork, and analytic outputs (21%) and data or spreadsheet work (14%) lead (appendix, Figure A.1).

Crossed with the use-case classifier introduced in January 2026 (work, personal or coursework), more than 80% of creative writing, guidance and recipe conversations are personal, while marketing content (80%), blogs and articles (81%) and database queries (82%) are mostly work; plans (44% work, 49% personal) and translation (42% and 44%) are split. Documents and reports lead work (20%) and coursework (21%) conversations, explanations (25%) and recommendations (22%) personal ones (Figure 2.2 and text).

Token use and the wage of the mapped occupation

For work conversations, each conversation's classified task is mapped to the occupation that usually performs it, and the typical (geometric mean) token count, including Claude's reasoning, is compared with that occupation's median wage (Figure 2.3, left). The relationship is positive and, in the report's words, "noisy." Marketing managers earn about twice what editors earn ($80 against $37 an hour) and their tasks use about 2.5 times as many tokens; pharmacists earn nearly three times what statistical assistants earn ($68 against $24) yet use about a twentieth as many. Token counts are not adjusted for the model serving the conversation (caption), though Chapter 2's footnote 3 says the relationship is similar when tokens are weighted by price. About 44% of the wage gradient is explained by the mix of outputs (apps and websites use more than three times the median conversation's tokens, explanations about a fifth).

Table 2.4 compares occupations in the top and bottom thirds by wage: 2.07 times the tokens per conversation, 1.53 times the turns, 1.34 times Claude's output per turn, 2.05 times the price-weighted compute cost, and extended thinking on in 34% of conversations against 31%. The authors read the rise in both Claude's output and users' turns as suggesting that "more production from Claude does not mean less from the user," a pattern they say "looks more labor-augmenting than labor-displacing" if the human stays involved.

Autonomy by surface

AI autonomy is Claude's rating, on a 1 to 5 scale, of how much decision-making is left to Claude. It is higher on Claude Code than on chat and Cowork for 26 of 31 output types; for scripts and code snippets the difference is 0.53 points, and across all conversations 0.37 (Figure 2.5 and text). About two thirds of the overall gap comes from the same outputs being made with more delegation, one third from the different mix of outputs. Model choice does not account for the gap: 54% of Claude Code sessions are served by Opus against 10% of chat and Cowork conversations, but among Sonnet conversations alone Claude Code still shows 0.26 points more autonomy. The exception is data and spreadsheets (3.09 on chat and Cowork against 2.74 on Claude Code), about 70% of which the report attributes to task mix (footnote 5). Across output types, mean autonomy and median token use correlate at r = 0.68 (appendix, Figure A.2).

Reading level

A classifier estimates the years of education needed to understand the prompt and the response (its prompt is in the March 2026 appendix). Prompt and response levels correlate at 0.87; responses run about one year above prompts on average, most for images and graphics (+2.6 years), games (+1.9) and apps and websites (+1.7), and near zero for blogs (−0.1), academic papers (+0.0) and email (+0.3) (Figure 2.6 and text). The report suggests part of the gap is register, terse prompts against polished replies.

Chapter 3: the first survey results

Sample

The Economic Index Survey invites a random sample of Claude users; its design is described in Announcing the Anthropic Economic Index Survey. For this report, each respondent's usage from mid-May to early June 2026 is linked to their answers through the privacy-preserving system, sampling up to 20 sessions per person across Claude.ai, Cowork and Claude Code and dropping respondents with fewer than five sessions. The linked sample is "about 9,700" respondents. The report gives no number invited, no response rate and no weighting, and states that the survey "is not representative of the general population" (section "Who responded"). Computer and mathematical occupations are about 30% of respondents against 4% of US employment, and management 23% against 7% of employment and 4% of sessions (Figure 3.1). Women are 12% of the linked sample.

Expected exposure to AI

Respondents chose from five bands the share of their work tasks AI could do "entirely on its own" today (reported exposure) and in 12 months (anticipated exposure). Close to 6 in 10 chose a higher band for next year, and over a third expect AI to do most or nearly all of their tasks next year (Figure 3.2). Across occupations, reported exposure correlates positively with Anthropic's earlier usage-based measure of observed exposure (Labor market impacts of AI: A new measure and early evidence) and with a published measure of theoretical exposure, and exceeds observed exposure; the gap between today and next year is about the same in highly and less exposed occupations (Figure 3.3). Footnote 4 notes that midpoint coding of the bands biases the slopes toward zero, so the authors compare them "qualitatively."

Reported exposure is about 10 percentage points lower in high-income countries, and about 10 points lower for people with at least 15 years of experience than for those in their first year (Figure 3.4). The report links the second pattern to tacit expertise, citing open-ended answers, and says "perceptions about future improvements" are "essentially uncorrelated" with GDP and experience. In Figure 3.4 the anticipated shares also fall with GDP and experience, roughly in parallel with the reported shares (read from the figure), so this statement appears to concern the expected increase rather than the anticipated level (this note's reading). Reported and anticipated exposure both rise with automation share, the share of a person's sessions classified as directive or feedback-loop interactions; the report offers two explanations, that delegating shows people what AI can do, or that those who already believe it can are readier to delegate.

Expectations about jobs

More than a third rated it likely or very likely that their own, a peer's, a junior or a senior colleague's responsibilities would change significantly in the next year. Ten percent rated losing their own job likely or very likely, which the report compares with a US layoff and discharge rate of about 13.4% a year (footnote 10); because respondents skew toward knowledge workers in stable jobs, it says this "may still indicate elevated perceived risk." Of those rating job loss likely, 38% mentioned AI when asked what drove their forecasts, an upper bound because the question covered job change and job loss together (footnote 11). Respondents were more worried about others than themselves, especially junior colleagues, for whom over a third put the chance of job loss above 60% (Figure 3.5).

On six dimensions of work (pay, job security, finding a new job, meaning, autonomy and human interaction), the share expecting a positive effect of AI rises with automation share for all six, most for pay and finding a job. Figure 3.6 prints the mean shares positive (pay 56%, security 42%, finding a job 52%, meaning 59%, autonomy 70%, human interaction 36%) and plots regression coefficients of roughly 3 to 5 percentage points per standard deviation (22 points) of automation share (read from the figure). On selection, enthusiasts delegating more, the authors write that they "can't rule this out entirely," but report that the estimates change little when controlling for tenure on Claude.ai, which they treat as a proxy for enthusiasm.

Most respondents report productivity gains in speed (86%), scope (82%) and quality (69%), 27% savings on services they would otherwise buy, 68% learning more and 57% skills made more valuable. The share reporting more valuable skills rises with automation share; the share reporting more learning is roughly flat (Figure 3.7). The report cautions that these are self-assessments and "do not rule out skill erosion."

Gender

Controlling for occupation (SOC minor groups), women's Claude Code share of sessions is 0.24 standard deviations (6.3 percentage points) lower than men's and their automation share 0.33 standard deviations (7.3 points) lower, with more active time on chat (Figure 3.8, with 95% intervals). Footnote 14 says the pattern holds when controlling for Claude Code share.

Hopes for the economy

A final open question asked respondents to "dream big" about the economy in ten years. A classifier, not further described, tagged responses with themes (definitions in appendix Table A.1). Over half hoped for collaboration with AI on meaningful work, just over half for automation of tedious work, and about a third for widely shared gains (Figure 3.9).

Comparisons with earlier reports

The report does not set its levels beside earlier reports'. Several shift between the March 2026 release (Claude.ai, 5 to 12 February 2026) and this one (chat and Cowork, April and May 2026): mean AI autonomy falls from 3.41 to 2.72 and 2.74, the mean estimated time for a human alone rises from 3.1 to 4.6 and 4.7 hours, and the automation share (directive plus feedback-loop conversations, excluding those whose pattern is "none", as the June release does) rises from about 46% to 49% (this note's reading and arithmetic from the two releases' global rows; with "none" kept in the denominator, 44% to 48%). The classifier, the transcript length shown to it, the O*NET version and the surfaces all changed between them, so these differences cannot be read as changes in use without a comparison on common data, which the report does not provide.

The prompts used

The measurements come from prompted Claude classifiers, not from an assistant whose behavior is studied. The appendix prints the O*NET prompt and the artifact prompt, which instructs: "Only choose an artifact label if it clearly describes what the exchange produced. If no label is a clear fit, choose none." The report does not test alternative wordings.

Limitations and open questions

  • Classifier validation. Every usage measure, and automation share in the survey analysis, is a Claude classification. The report names no classifier model and reports no agreement with human coders; the appendix's WildChat examples are illustrations.
  • A survey of users, linked to usage. The sample is Claude users, filtered to those with at least five sessions, with no response rate or weighting reported. Associations between usage and expectations are cross-sectional; the authors name selection as a possible explanation and test it only through a tenure control.
  • Usage share is not impact. The token-wage gradient is computed from occupations mapped to tasks, not from users' jobs.
  • Sample windows differ by chapter: parts of April to June 2026 (Chapter 1), 10 April to 10 June (Chapter 2), mid-May to early June (survey linkage).
  • Interest. This is a model developer's report on use of its own products; the survey is run inside those products.

Research questions

  1. How much of the change in primitive levels since February comes from the new pipeline? Running the old and new classifiers on the same conversations would separate method from use.
  2. Is the weekend rise in personal use a change in what the same people do, or in who uses Claude? A per-account panel, within the privacy limits, would distinguish the two.
  3. Does the token-wage gradient reflect the value of the work or the models and prices chosen? The report notes model mix is unadjusted in the plot; regressions with model fixed effects and output controls would test what remains.
  4. Does automation lead to optimism, or optimism to automation? The survey is monthly; following the same respondents over time, or an encouragement design, would test the order.
  5. How accurate is the artifact classifier? Human coding of a sample, including the "none" category, would give agreement rates by label.
  6. Who answers the survey? Comparing invited and responding users on usage measures would show the direction of nonresponse bias.
  7. Does the gender gap in delegation persist within detailed occupations and experience levels? Finer controls and more months of data would test it.

How it connects

References