Anthropic Economic Index March 2026 Report: Learning Curves

TL;DR

Anthropic’s March 2026 Economic Index report finds that Claude usage diversified across tasks and wage levels, while experienced users achieve higher success rates and tackle higher‑value work, suggesting learning‑by‑doing and emerging skill‑biased economic effects.

Overview of New Findings

• Task diversification: The top 10 tasks on Claude.ai fell from 24 % of traffic in November 2025 to 19 % in February 2026, indicating a broader spread of use cases.
• Shift toward lower‑wage tasks: Average task value on Claude.ai dropped from $49.3 to $47.9 per hour, driven by more personal queries (sports, product comparisons, home maintenance) and migration of coding work to the API.
• Geographic concentration: The top 20 countries now account for 48 % of per‑capita usage (up from 45 %). Within the United States, usage is slowly converging across states, with the top 5 states’ share falling from 30 % to 24 %.
• Learning curves: Users with ≥ 6 months tenure use Claude for higher‑education tasks, have 10 % higher conversation success rates, and allocate 10 % fewer conversations to personal use, even after controlling for task, country, and model.

Diversification of Use Cases on Claude.ai

• Sampling method: Anthropic’s privacy‑preserving system aggregated 1 M conversations from Claude.ai (web) and the first‑party API.
• Coding migration: Coding tasks remain the largest category (35 % of Claude.ai conversations) but have shifted to the API, where they are split into many smaller tasks, keeping API task concentration flat.
• Work vs. personal mix: Coursework fell from 19 % to 12 % of conversations, while personal use rose from 35 % to 42 %.
• Collaboration modes: Augmentation (validation, learning) increased slightly on Claude.ai, while automation decreased sharply on the API.

Model Selection Aligned with Task Value

• Opus preference: Paying Claude.ai users select the Opus model 4 pp more than average for coding tasks and 7 pp less for tutoring. API users show an even stronger bias.
• Wage correlation: For Claude.ai, each $10 increase in estimated hourly wage of a task raises Opus usage by 1.5 pp; for the API the increase is 2.8 pp.
• Occupational breakdown: 34 % of Software Developer tasks use Opus versus 12 % of Tutor tasks on Claude.ai.

Learning Curves and Tenure Effects

• Definition of tenure groups: High‑tenure users signed up ≥ 6 months before the data pull; low‑tenure users are newer.
• Behavioral differences:
  • High‑tenure users are 7 pp more likely to use Claude for work.
  • Their conversations involve 1 year more of required human education on average.
  • Personal‑use share drops from 44 % (new users) to 38 % (one‑year‑old users).
  • Success rate is 5 pp higher in a simple regression; after adding task fixed effects it remains 3 pp higher, and 4 pp higher with full controls.
• Interpretation: The evidence supports learning‑by‑doing—users become better at prompting and task framing over time—though cohort and survivorship biases cannot be fully ruled out.

Geographic Convergence and Divergence

• US states: Per‑capita usage is spreading; the Gini coefficient fell, but convergence is now projected to take 5–9 years (previously 2–5.8 years).
• Countries: Usage concentration increased; the top 20 countries’ share rose from 45 % to 48 %.

Emerging Automation Patterns in the API

• Automation rise: API workflows are increasingly directive, reducing human‑in‑the‑loop steps.
• Fast‑growing use cases (share doubled since the previous report):
  • Business sales & outreach automation (lead qualification, cold‑email drafting).
  • Automated trading & market operations (monitoring positions, investment suggestions).

Implications for the Labor Market

• Skill‑biased technological change: Early adopters with high‑skill, high‑wage tasks achieve higher AI success rates, potentially widening wage gaps.
• Potential feedback loop: Users who master Claude may extract more value, reinforcing their advantage, while later adopters face lower‑value tasks.
• Policy relevance: Tracking these dynamics early gives researchers and policymakers time to design interventions that mitigate inequality.

Limitations and Future Work

• Data scope: The report covers only a one‑week window (Feb 5‑12 2026) and may not capture longer‑term trends.
• Survivorship bias: Users who stopped using Claude are not observed, possibly inflating success estimates for high‑tenure cohorts.
• Cohort effects: Early adopters may differ systematically from later users; future analyses will aim to separate cohort from experience effects.

Data and Citation

• Dataset: Available at https://huggingface.co/datasets/Anthropic/EconomicIndex.
• Citation:

@online{anthropic2026aeiv5,
  author = {Maxim Massenkoff and Eva Lyubich and Peter McCrory and Ruth Appel and Ryan Heller},
  title = {Anthropic Economic Index report: Learning curves},
  date = {2026-03-24},
  year = {2026},
  url = {https://www.anthropic.com/research/economic-index-march-2026-report}
}