Anthropic Economic Index: cadences of Claude use | Keryc
A year ago, using Claude was mostly a conversation between a person and an assistant. Have you noticed how that changed? With Claude Code and Cowork growing, many sessions are now long, agentic tasks — it’s no longer enough to read a chat transcript to understand the economic impact. Anthropic updates its Economic Index with more frequent sampling, new classifiers, and a usage-linked survey to show how AI embeds itself into the rhythms of work and daily life.
What changed in this report
Hourly sampling instead of weekly samples: now you can see patterns by the hour and by day.
A new classifier tags the main artifact of each conversation (document, explanation, code, etc.).
More granular data: they separate chat and Cowork (together, “Claude conversations”) from the 1P API, aggregated monthly.
Private telemetry: transcripts are only processed by another Claude instance for classification, and cells with few observations are filtered out.
These changes let you analyze daily cadences, the types of outputs Claude generates, and how people perceive its impact on their work.
Cadences: workweek, daily rhythms and spikes (for example, tax day)
The share of personal conversations in chat and Cowork rises from ~35% on weekdays to almost 50% on weekends. What shifts? From emails and presentations to emotional support, medical questions, and investment advice.
Clear hourly rhythms: people ask for news at 7 a.m., recipes at 6 p.m. (2.3x the average), and sleep advice around 3 a.m.
Important dates show up in the data: tax-related requests spiked eightfold right before the U.S. deadline on April 14.
Weekends also show more conversations about starting businesses, while job applications drop. And when people work outside typical hours, the tasks tend to map to higher-paid occupations.
Artifacts: what Claude produces and how it’s used
Anthropic classified each conversation into >30 artifact categories. Key result: 93% of conversations produce a visible artifact.
Most common: explanations (17%), documents and reports (15%), and guides or recommendations (11%).
Work conversations tend to generate documents, email drafts, analyses and summaries. Personal conversations produce more explanations and recommendations.
Some categories are almost exclusively personal (creative, recipes, guidance), while marketing, blogs and database queries are mostly work-related.
This helps map not just how much Claude is used, but what you get out of each interaction.
Computational cost and value: tokens as a proxy
Anthropic uses tokens (text processed and generated, including internal reasoning) as a measure of a conversation’s “cost.” They use geometric means because consumption is highly skewed.
There’s a positive relationship between tokens per conversation and the average salary of the mapped occupation: higher-paid occupations tend to use more tokens.
Example: conversations mapped to marketing managers consume ~2.5x the tokens of those for editors, matching salary differences.
Complex artifacts (apps, sites) use far more tokens than simple explanations.
The economic implication: more tokens usually correspond to higher-value outputs, and in higher-paid occupations both Claude and the user produce more during the session — suggesting productivity gains rather than simple replacement.
Autonomy: how much Claude decides on its own
Measured on a 1–5 scale (none to extreme). Tasks like translations or calculations are low autonomy; building apps, games or presentations is high autonomy.
Claude Code consistently shows more delegation than chat or Cowork: an average difference of 0.37 points. Part of that is users delegating more in Code and part is the different task mix.
It’s not just the model: although Code uses Opus more often (54% vs 10% in chat/Cowork), the difference remains between sessions served by the same model (for example Sonnet), suggesting the product changes behavior.
Autonomy and tokens move together (correlation r = 0.68): more autonomy implies a more computationally costly dialogue.
Reading level: Claude raises the register
Anthropic estimated years of education needed to understand the prompt and the response. Important findings:
On average, Claude’s response is ~1 year above the prompt in educational level.
Larger gaps appear when asked to build something: images/graphics +2.6 years, games +1.9, apps and sites +1.7.
For writing aimed at an audience (blogs, emails, papers) the gap is close to zero.
That could reflect prompts being concise and outputs coming in polished prose, but it also raises questions about accessibility and adapting language for different audiences. How do you make sure outputs remain approachable?
Perceptions: the usage-linked survey (n ≈ 9,700)
Anthropic launched the Anthropic Economic Index Survey in April 2026. They linked responses to up to 20 sessions per user in a sampling period using privacy-preserving methods.
Most expect rapid progress in AI capabilities over the next year; about 6 in 10 anticipate greater capability in 12 months.
More than a third think AI could do most or nearly all of their work within a year.
An interesting tension: users who delegate more (more automated sessions) are the most optimistic about pay, job security, and meaning of work.
Early-career workers are more worried about displacement, while those who delegate report more productivity gains and learning.
How they measure automation and collaboration modes
The mode classifier distinguishes: Directive (you delegate execution), Feedback Loop, Task Iteration, Learning and Validation. The “automation share” is the fraction of sessions with directive or feedback-loop patterns.
The report shows reported and anticipated exposure rise with automation share: delegating is informative about capabilities, or enthusiasts simply delegate more.
Employment, risks and inequalities
10% of respondents said losing their job was likely or very likely within 12 months; many attribute job-loss risk to AI.
Greater concern for junior colleagues and for workers in lower-income countries.
Gender difference: in the linked sample, women (12% of the sample) use Claude Code less, have lower automation share, and work more iteratively with Claude.
Methodology, limits and privacy
Analysis period for chapters: mainly April 10 to June 10, 2026.
Data include chat and Cowork (consumer Free/Pro/Max) and 1P API separately; Claude Code is analyzed where relevant.
Privacy: private classifiers read transcripts with another Claude instance and scarce cells are filtered to avoid reidentification.
Tokens are measured and summarized with geometric means due to heavy skew in consumption.
What this leaves us (and what to ask)
This report confirms what we already suspected: AI isn’t just a point tool — it’s part of daily work and life rhythms. It produces concrete artifacts, demands more compute when the work is worth more, and receives different levels of delegation depending on the surface and the task.
Open questions that matter: how do we ensure higher-reading outputs are accessible to broad audiences? How do we measure erosion of real skills versus self-reported learning? What social and training policies do we need if productivity benefits concentrate in certain groups?
The good news is that data plus the survey give a richer picture: not just how much Claude is used, but what’s produced and how people perceive it. That’s key to shaping informed policy and business practices.
