Anthropic Economic Index September 2025 report

TL;DR

AI adoption is heavily concentrated in wealthy countries and a few U.S. states, with enterprise API usage focused on coding and automation, suggesting that productivity gains may accrue primarily to already‑rich economies.

Key Findings Overview

Geographic concentration: The Anthropic AI Usage Index (AUI) shows that high‑income countries use Claude AI many times per working‑age capita more than low‑income nations (e.g., Israel 7×, Singapore 4.6×, India 0.27×).
Task diversification: In high‑adoption regions, usage spreads beyond coding to education, science, and business tasks, while low‑adoption regions remain coding‑centric.
Collaboration mode shift: As per‑capita usage rises, users move from automation‑dominant “directive” interactions to more collaborative “augmentation” patterns.
Enterprise API deployment: 77% of API traffic is automation‑dominant, with coding and office/admin tasks over‑represented; cost sensitivity is weak, and contextual data availability limits complex deployments.

1. Changing Claude.ai Usage Over Time

Conclusion: Claude’s usage has shifted toward knowledge‑intensive tasks and more autonomous interactions, reflecting both model improvements and new product features.

Education and science growth: Educational task share rose from 9.3% to 12.4%; scientific tasks grew from 6.3% to 7.2%.
Coding still dominates: Coding accounts for 36% of all sampled conversations.
Directive automation up: “Directive” conversations increased from 27% to 39%, making automation the majority interaction mode for the first time.
Feature‑driven task changes: Web‑search release (Mar 2025) boosted “search electronic sources” from 0.03% to 0.49%; Research mode (Apr 2025) raised “internet‑based research” from 0.003% to 0.27%; instructional‑material creation rose six‑fold (0.2% → 1.5%).
Code creation vs. debugging: Code‑creation tasks grew by 4.5 pp (4.1% → 8.6%); debugging fell by 2.9 pp (16.1% → 13.3%).

"The jump in directive usage could also signal growing confidence in delegating complete tasks to AI, a form of learning‑by‑doing." – Anthropic report

2. Geographic Diffusion of AI Adoption

Conclusion: AI usage per capita correlates strongly with national income, leading to a global divide where high‑income economies reap disproportionate benefits.

2.1 Global Usage Patterns

U.S. dominance: The United States accounts for 21.6% of total Claude usage.
AUI correlation with income: Countries with higher GDP per working‑age capita show higher AUI values (e.g., Israel 7×, Singapore 4.6×, Canada 2.9×). Emerging economies such as Indonesia (0.36×), India (0.27×), and Nigeria (0.2×) lag far behind.
Task mix differences: Low‑AUI countries allocate >50% of usage to coding, whereas high‑AUI regions spread usage across education, science, and business.
Automation vs. augmentation: Controlling for task mix, high‑AUI countries exhibit more augmentation (collaborative) use, while low‑AUI countries favor automation.

2.2 United States State‑Level Insights

Per‑capita leaders: District of Columbia leads with an AUI of 3.82, closely followed by Utah (3.78), surpassing California (2.13).
Economic correlation: Each 1% increase in state GDP per capita is associated with a 1.8% increase in the AI Usage Index, though income explains less than half of cross‑state variation.
Specialized usage: DC shows over‑representation in document editing, information provision, and job‑application assistance; California emphasizes IT and digital‑marketing tasks; Florida leans toward business advice and fitness‑related queries.

3. Enterprise API Deployment of Claude

Conclusion: Early enterprise adopters use Claude primarily for high‑value, automation‑friendly tasks such as software development, with limited price sensitivity and a strong dependence on readily available contextual data.

3.1 API Usage Landscape

Task concentration: ~44% of API traffic maps to Computer & Mathematical occupations; Office/Admin tasks account for ~10%.
Automation dominance: 77% of API transcripts exhibit automation patterns (full task delegation) versus ~12% augmentation.
Cost vs. usage: Higher‑cost tasks (e.g., advanced coding) have higher usage shares; a 1% cost increase reduces usage by only 0.29% after controlling for task characteristics, indicating weak price elasticity.

3.2 Contextual Information as a Bottleneck

Input‑output relationship: Across tasks, a 1% increase in input token length yields a 0.38% increase in output token length, showing diminishing returns on longer contexts.
Complex task constraints: Tasks requiring extensive context (e.g., sales‑strategy generation) may face adoption hurdles if firms lack centralized data repositories.

"Access to appropriate contextual information is needed for high‑impact deployments of AI in complex domains." – Anthropic report

4. Economic Implications of Uneven Adoption

Conclusion: The current concentration of AI usage threatens to amplify global inequality and reshape labor markets, benefiting high‑skill workers while potentially displacing lower‑skill roles.

Growth convergence risk: Historical parallels with electricity and the internet suggest that early concentration can lead to lasting divergence; AI’s productivity gains may therefore accrue mainly to rich regions.
Labor market dynamics: Automation‑heavy adoption could displace workers in routine coding and data‑entry tasks, whereas workers with tacit organizational knowledge may see higher demand as complements to AI.
Policy considerations: To avoid deepening digital divides, policymakers should address infrastructure gaps, promote AI literacy, and consider regulations that encourage broader, inclusive AI diffusion.

5. Open Data for Independent Research

Anthropic has released the underlying task‑level usage data for Claude.ai and 1P API traffic (geographic breakdowns currently only for Claude.ai). The dataset includes:

O*NET‑mapped task classifications
Bottom‑up request taxonomies
Collaboration‑mode breakdowns
Methodology documentation

Researchers are invited to explore questions such as local labor‑market impacts, determinants of cross‑country AI adoption, price sensitivity of enterprise deployment, and the role of contextual data in shaping AI effectiveness.

6. Concluding Remarks

Anthropic’s September 2025 Economic Index highlights that AI adoption is proceeding at unprecedented speed but remains highly uneven across regions and enterprise use cases. The concentration of usage in high‑income economies and automation‑centric enterprise deployments suggests that productivity gains may be unevenly distributed, potentially widening existing economic gaps. Continued monitoring, open data sharing, and proactive policy interventions will be essential to steer AI’s transformative potential toward inclusive growth.