Anthropic Economic Index Cadences Report – June 2026

TL;DR

Anthropic’s June 2026 Economic Index shows that Claude’s usage now mirrors hourly work rhythms, that higher‑wage occupations generate larger‑token (more compute‑intensive) artifacts, and that users who delegate tasks to Claude are the most optimistic about AI’s future impact on pay, job security, and skill value.

1. New Measurement Cadence Captures Hourly Work Patterns

Conclusion: Hourly sampling reveals that Claude usage spikes for work‑related queries during the workday and for personal queries on weekends, confirming that AI adoption follows real‑world labor rhythms.

Data are now sampled hourly rather than in seven‑day windows, enabling fine‑grained analysis of daily cycles.
Personal‑use conversations rise from ~35 % on weekdays to just under 50 % on weekends (Figure 1.1).
Hourly peaks include news requests at 7 am, recipe requests at 6 pm (2.3× average), and sleep advice around 5 am (Figure 1.2).
Tax‑related queries surge eight‑fold on April 14‑15, aligning with the U.S. filing deadline (Figure 1.4).
Night‑time and weekend work conversations are disproportionately from higher‑wage occupations, even after removing computer‑math tasks.

2. Artifact Classification Shows What Claude Produces

Conclusion: Over 90 % of Claude conversations generate a concrete artifact; the most common are explanations (17 %), documents/reports (15 %), and guidance (11 %).

Artifacts are grouped into >30 categories; the distribution is shown in Figure 2.1.
Work‑related artifacts dominate documents, explanations, and email drafts, while personal artifacts are dominated by explanations (25 %) and recommendations (22 %).
Certain artifact types are almost always personal (e.g., creative writing, recipes), whereas others are predominantly work‑related (e.g., marketing content, database queries).

3. Compute Consumption Correlates with Economic Value

Conclusion: Conversations mapped to higher‑wage occupations consume more tokens, and token usage is strongly linked to the economic value of the output.

Median token count rises with median occupational wage (log‑log relationship, Figure 2.3 left).
Building apps consumes >3× the median token count, while simple explanations consume ≈0.2×.
About 44 % of the wage‑gradient in token usage is explained by the mix of artifact types (higher‑wage jobs produce more compute‑intensive artifacts).
Higher‑wage conversations also feature 1.34× more Claude output per turn and 1.53× more user turns, indicating labor‑augmentation rather than displacement (Table 2.4).

4. AI Autonomy Varies by Surface and Output

Conclusion: Claude Code sessions exhibit higher AI autonomy than chat or Cowork, even when the same model (e.g., Sonnet) is used.

Autonomy is rated on a 1–5 scale; Claude Code scores 0.37 points higher on average (Figure 2.5).
The gap stems mainly from greater delegation in Claude Code (≈2/3 of the difference) and from a different output mix (≈1/3).
High‑autonomy outputs (e.g., app or website creation) also have the highest token consumption (correlation r = 0.68).

5. Claude’s Responses Are Generally More Complex Than Prompts

Conclusion: Across artifact types, Claude’s replies are on average one education‑year higher than the user prompt, with the largest gaps for design‑heavy tasks (e.g., graphics + 2.6 years).

Reading‑level estimates use years of education required for comprehension (Figure 2.6).
Academic‑paper outputs require >16 years; recipes require <10 years.
The prompt‑response gap is near zero for audience‑facing writing (blogs, emails).

6. Survey Insights: Expectations, Optimism, and Demographics

Conclusion: Surveyed Claude users (≈9.7 k respondents) overwhelmingly expect AI capabilities to expand in the next year, and those who automate more tasks are the most optimistic about pay, job security, and skill value.

Respondents are skewed toward computer‑mathematical (≈30 %) and management (≈23 %) occupations, far above their shares in U.S. employment.
60 % anticipate a higher share of their tasks being AI‑automatable in 12 months; >33 % expect AI to handle most or all of their work.
Reported exposure aligns with observed and theoretical exposure, but anticipated exposure is uniform across occupations.
Higher automation share correlates positively with optimism on all six job‑quality dimensions (Figure 3.6).
Women (12 % of sample) use Claude less for work and automation, and more iteratively, even after controlling for occupation (Figure 3.8).

7. What People Hope for in an AI‑Transformed Economy

Conclusion: The most common aspirations are collaborative AI augmentation of meaningful work, automation of drudgery to free time, and broad sharing of AI‑generated economic gains.

Over 50 % of open‑ended responses mention AI‑human collaboration on meaningful tasks.
A similar share desire automation of tedious work for more leisure.
About one‑third hope that AI’s productivity gains will be widely distributed.

8. Implications for Researchers and Policymakers

Conclusion: Fine‑grained usage data and artifact classification provide early indicators of AI’s economic diffusion, while survey results highlight a gap between observed exposure and user expectations that could shape labor‑market policy.

Hourly telemetry enables real‑time monitoring of AI adoption cycles and can inform workforce‑planning tools.
The token‑wage relationship suggests compute cost can serve as a proxy for the economic value of AI‑generated work.
Optimism among high‑automation users may drive faster adoption, but the uniform expectation of rapid capability growth across occupations warrants caution in forecasting sector‑specific impacts.
Gender‑based usage differences indicate a need for inclusive design to ensure equitable access to high‑autonomy AI tools.

The full methodological appendix and data visualizations are available in the original report’s PDF.