Anthropic Economic Index: Tracking AI’s role in the US and global economy

economic impactAI adoptionautomation and augmentationgeographyopen data

Anthropic's summary post for the September 2025 Economic Index report presents its results for a general audience (uneven Claude use across countries and US states, a rise in directive automation from 27% to 39%, and heavier automation by API business customers) and announces an interactive website and open dataset.

Anthropic · 2025-09-15 · original

This post accompanies the third Anthropic Economic Index report, Anthropic Economic Index report: Uneven geographic and enterprise AI adoption, published the same day. It summarizes the report in a much shorter form for a general reader, opening with the uses of Claude that are most overrepresented in particular places: travel planning in Hawaii, scientific research in Massachusetts, building web applications in India. It also launches an interactive website where readers can look up Claude.ai use by US state and occupation, and points to the open dataset on Hugging Face. The full report holds the methods, footnotes and caveats; this note records what the post says and where its presentation differs, and defers to the report note for definitions and detail.

What the post summarizes

Geography

  • Countries. The United States uses Claude far more than any other country, followed by India, Brazil, Japan and South Korea "each with similar shares." To adjust for population, the post introduces the Anthropic AI Usage Index (AUI), a country's share of Claude.ai use divided by its share of the working-age population. Small, technologically advanced countries such as Israel and Singapore lead, and a 1% higher GDP per capita is associated with a 0.7% higher AUI (section "Across countries"). The post raises the possibility that, as with electrification and the combustion engine, gains could concentrate in richer countries.
  • Language use in Brazil. Users in Brazil use Claude for translation and language learning "about six times more than the global average" (introduction).
  • US states. Within the US, a 1% higher GDP per capita is associated with a 1.8% higher population-adjusted use, but income explains less of the variation than across countries. The District of Columbia has the highest AUI (3.82), with document editing and information search among its overrepresented uses; coding is especially common in California (third-highest AUI) and finance in New York (fourth). In Hawaii, tourism-related requests occur at twice the rate of the rest of the country (section "Patterns within the United States"). Footnote 1 notes that part of Utah's activity, second in AUI, appeared linked to coordinated abuse, which robustness checks suggested was not driving the results.
  • Tasks. Computer and mathematical uses have been around 37 to 40% of conversations since December 2024. The post reports educational instruction rising "by more than 40 percent (from 9% to 13% of all conversations)," physical and social sciences from 6% to 8%, management falling from 5% to 3%, and business and financial operations halving from 6% to 3% (section "Tasks"). The report states these shares in more than one rounding: its introduction gives 9.3% to 12.4% for education and 6.3% to 7.2% for science, its Chapter 1 text 9% to 12% and 6% to 7%, and the labels on its Figure 1.1 end at 12.7% and 7.4%, which the post's 13% and 8% match.
  • Collaboration. Directive conversations rose from 27% to 39%, and "for the first time, automation (49.1%) has become more common than augmentation (47%) overall" (section "Patterns of interaction"). The post suggests improved models as one cause, noting that the latest model in December 2024 was Sonnet 3.6. The report adds a caveat the post omits: the classifier model also changed between reports, and re-running the latest data with the older classifier gave a lower automation share (45% rather than 49%), though directive use still rose significantly (report, Chapter 1, footnote 4).
  • Collaboration by country. Controlling for task mix, "a 1% increase in population-adjusted use of Claude is correlated with a roughly 3% reduction in automation"; the post says the reason is not yet known.

Business use

API customers, mostly businesses and developers, were sampled for the first time: 1 million transcripts from August 2025, from customers accounting for roughly half of 1P API usage (footnote 3). 44% of API traffic maps to computer and mathematical tasks against 36% on Claude.ai, about 5% to developing and evaluating AI systems, educational tasks 4% against 12%, and arts and entertainment 5% against 8%. 77% of API conversations show automation patterns, mostly directive, and 12% augmentation, while Claude.ai is close to even. Higher-cost task categories see more use, which the post reads as capability and value mattering more to businesses than cost (section "Businesses").

The post's conclusion

The post summarizes adoption as "remarkably uneven": higher-income countries use Claude more, more collaboratively and for a wider range of tasks; state-level use follows local industries; and businesses grant more autonomy than consumers. It singles out the growth of directive automation over nine months and says the authors will watch whether users' choices settle as models improve.

Differences from the full report

  • Rounding. The post rounds figures that the report gives more precisely, often in its charts rather than its text: the education and science shares (above), "about six times" for translation and language learning in Brazil (6.4x in report Figure 2.8), and "roughly 3%" for the automation reduction (a partial-regression coefficient of magnitude 3.112 in report Figure 2.11).
  • Caveats. The post omits the classifier change (report, Chapter 1, footnote 4), the regression controlling for task characteristics that finds a small negative cost elasticity (report, Figure 3.9), and the context-length analysis (report, Chapter 3). Readers relying on the post should check these in the report.
  • No principles or prompts. Like the report, the post contains no constitution or behavior-shaping text; its results come from Claude classifying conversations through the privacy-preserving system Clio, which the post links for method.

Research questions

The questions raised by these results are set out in the report note. Two concern the post specifically:

  1. How do summary posts change the reading of a report? The post omits the classifier caveat on the automation trend. Comparing how press coverage and later citations describe the trend, depending on whether they cite the post or the report, would show which version circulates.
  2. How is the interactive data used? The post invites readers to explore state and occupation data. Download and query statistics for the website and dataset, and the studies that cite them, would show whether the open data serve the independent research the report hopes for.

How it connects

References

  • Anthropic. (2025, September 15). Anthropic Economic Index: Tracking AI's role in the US and global economy. https://www.anthropic.com/research/economic-index-geography
  • Appel, R., McCrory, P., Tamkin, A., Stern, M., McCain, M., and Neylon, T. (2025, September 15). Anthropic Economic Index report: Uneven geographic and enterprise AI adoption. https://www.anthropic.com/research/anthropic-economic-index-september-2025-report
  • Anthropic Economic Index dataset: https://huggingface.co/datasets/Anthropic/EconomicIndex