Anthropic Economic Index

Confidence 0.85 · 5 sources · last confirmed 2026-06-17

A research initiative by Anthropic that measures real-world AI use through privacy-preserving analysis of Claude conversations on Claude.ai (consumer) and the 1P API (enterprise), and — increasingly — Claude Code agent sessions. Recurring report cadence — through June 2026 the wiki holds the 4th and 5th editions, the March labor-impacts note, and the June agentic-coding report.

Stated mission (per Anthropic): provide ongoing, empirical measurement of how AI is changing tasks, occupations, and the labor market.

What it tracks

DomainMapped to wiki concept
Task speedup, success, complexitygenerative-ai
Automation vs. augmentation shareautomation-vs-augmentation
Aggregate productivity impactai-employment-effects
Task-composition shiftai-deskilling
Cross-country adoptionenterprise-ai-adoption
Task-horizon time scalingai-benchmarks

Reports

ReportSample periodStatus in this wiki
FirstJanuary 2025Not separately ingested; numbers cited
Second(early 2025)Not separately ingested
ThirdAugust 2025Not separately ingested; numbers carried over
FourthNovember 2025Ingested — introduces “economic primitives” framework
Fifth — Learning curvesFebruary 5–12, 2026Ingested — model selection matches task value; high-tenure users have ~3-4 pp higher success after controls; skill-biased technological change framing
Labor-market impacts (Massenkoff & McCrory)March 5, 2026Ingested — the analytic/labor-impact branch of the AEI: introduces observed exposure (theoretical capability × usage, weighting automated/work-related uses), validates against BLS 2024–2034 projections, finds no systematic unemployment effect yet but a ~14% young-worker hiring slowdown into exposed occupations. The methodological primary behind the wiki’s “observed exposure” claims.
Agentic coding and persistent returns to expertise (Hitzig, Massenkoff, Lyubich, Heller & McCrory)October 2025 – April 2026 (~400,000 Claude Code sessions)Ingested — the agentic-coding branch of the AEI: the planning/execution division of labor (people make ~70% of planning decisions, Claude ~80% of execution), persistent returns to domain expertise (not coding skill; every major occupation succeeds within ~7 pts of software engineers), competence-captures-most-of-the-benefit, and the 7-month composition shift (fixing 33%→19%; writing/analysis ~doubled; task value +27%).

Economic primitives (introduced in fourth report)

Five measurements per conversation, derived by Claude classifying its own conversation samples:

  1. Task complexity — human time required without AI; whether multiple tasks were handled within one conversation.
  2. Human and AI skill level — years of education needed to understand prompts and Claude’s responses.
  3. Use case — work / education / personal.
  4. AI autonomy — degree of user delegation, from collaboration to fully directive.
  5. Task success — Claude’s own assessment of whether the task was completed.

See the fourth-report source page for definitions and applications.

Methodology notes

  • Privacy-preserving — random samples (typically 1M conversations on Claude.ai + 1M API transcripts).
  • Tasks mapped to O*NET taxonomy. O*NET vintage shifted between the 4th and 5th reports (4th used 2010 vintage; 5th uses 2019). Year-over-year comparisons of task-share need this caveat.
  • Models change report-to-report — fourth report uses Claude Sonnet 4.5 predominantly; the fifth uses Claude Opus 4.5 / 4.6 in addition. This affects comparability across editions.

Cited by external research

Open questions

  • Earlier reports (1st through 3rd) are referenced indirectly through the fourth-report carry-over data; first-party ingestion of any prior report would clarify the longitudinal methodology.