Anthropic's latest report, "Human Economy Index: Tracking the Role of AI in the US and Global Economy,"

Human Economic Index: TrackingAIThe role of the United States in the global economy is the main research topic.AI(in particularClaudeThe report examines usage globally and in various US states. It found that in the United States...ClaudeUsage far exceeds that of other countries, with India, Brazil, Japan, and South Korea following closely behind. At the national level,ClaudeThe use of [this technology] is positively correlated with GDP per capita, and high-income countries tend to use it more.ClaudeEngaging in knowledge work, not just simple programming. In the United States, various states...AIUsage varies depending on economic structure; for example, Washington D.C. is more often used for document editing, while California is more geared towards programming tasks. The report found that over time...ClaudeofautomaticThe proportion of tasks completed increased from 27% to 39%, indicating that users are more interested in...AITrust levels are increasing.

Travel planning in Hawaii, scientific research in Massachusetts, and web application development in India—on the surface, these three seem unrelated. In fact,ClaudeIts applications are most widespread in these places..

This does not mean that these areClaudemostPopularApplications: Software engineering still holds an absolute dominant position in almost every state and country globally. The so-called "Most widespreadThis refers to the fact that users in Massachusetts are far more likely to seek research assistance than those in other regions, and that Brazilian users are particularly enthusiastic about language-related tasks—they use...ClaudeThey engage in translation and language learning six times more frequently than the global average.

These findings come from our third issue of […].Anthropic Economic IndexIn this report, we further track...AIEarly trends in how technology is reshaping the work and economic landscape, with a focus on analysis.ClaudeDifferences in application across different dimensions:

Comparison of US statesInitial detailed assessmentAIInterstate differences in usage rates. It was found that each state's economic structure directly affects per capita usage.ClaudeUsage volume, and the states with the highest usage rates are not primarily focused on programming tasks.
Cross-border comparisonNew analysis shows that national income levels andClaudeUsage intensity is highly correlated; users in low-income countries tend to use...AIautomaticChemical work tasks.
Evolution of Time:contrastDecember 2024 - January 2025andFebruary-March 2025Data reveals that "command-type"automaticThe proportion of "transformation" tasks surged from 27% to 39%, indicating thatAIResponsibilities undertaken (and user trust)fastpromote.
Enterprise usersAdded Anthropic first-party API client anonymization data (overlay)ClaudeUsers outside of .ai), revealing enterprise usage patterns for the first time: API usersautomaticThe significantly higher propensity for task-oriented jobs compared to consumers suggests that the labor market may be about to undergo major changes.

The following is a summary of the report's core elements. We also launched [a similar service/product/service] simultaneously.Interactive websiteUsers can independently search for information in all states across the US and for all tracked occupations.Claude.ai usage trends, observe local or related industriesAIApplication scenarios. For in-depth analysis, we offer the current and past economic index datasets in their entirety.

This index adds a geographical dimension of data, revealing...ClaudeApplication characteristics in countries around the world and in various states of the United States.

cross-border model

USAClaudeUsage volume is far ahead, with India ranking second, and Brazil, Japan, and South Korea having similar usage shares.

Considering the differences in population size among countries, we use the “Anthropic AI Usage Index (AUI)” for correction analysis: AUI>1 indicates that the per capita usage in that country is higher than the expected level of its working-age population.

Data shows that technologically advanced small economies (such as Israel and Singapore) are in Claude In terms of adoption rate, it leads the working-age population. Income is likely a key factor: GDP per capita and the AUI index show a strong correlation (for every 1% increase in GDP per capita, AUI increases by 0.7%). This makes sense—ClaudeCountries with high usage typically possess well-developed internet infrastructure and knowledge-based, industry-driven economies. This also raises concerns about economic polarization: history shows that while general-purpose technologies such as electrification and the internal combustion engine have driven economic growth, they have also exacerbated global disparities in living standards.AIThe benefits are concentrated in high-income countries, which may lead to a recurrence of similar economic divergence effects.

US Interstate Model

GDP per capita andClaudeThe positive correlation between usage and per capita usage also holds true at the US state level, and the correlation is even stronger (for every 1% increase in GDP per capita, per capita usage increases by 1.8%). However, the explanatory power of income is lower than that of cross-national data, indicating the presence of other influencing factors.

We speculate that the economic structure of each state is the main factor. The highest AUI in the United States is in the District of Columbia (3.82), among which... Claude The most frequent uses are editing documents and searching for information, as well as other tasks related to knowledge work in Washington, D.C. Similarly, coding-related tasks are more common in California (the state ranked third overall by AUI).

Even in states with low usage rates like Hawaii, travel-related tasks are used twice as frequently as the national average. More data can be found through...Interactive websiteexplore.

Claude usage trends

We have been tracking how people use it since December 2024. ClaudeWe use a privacy-protecting methodClassification methodsAnonymous conversation logs are categorized according to task groups defined by O*NET (the US government database used to classify jobs and related tasks).In this way, we can analyze the allocations people have made since last year. Claude How have the tasks changed, and how have people chosen to cooperate (their choices regarding...)? Claude How much supervision and investment has been put into the work, and how have these changes occurred?

Task type

Since December 2024,Claude Computer and mathematical applications dominate our category, accounting for approximately 37-40% of the conversation.

But the situation has changed significantly. Over the past nine months, we've seen continued growth in "knowledge-intensive" areas. For example, educational guidance tasks have increased by over 40% (rising from 9% to 13% of all conversations), and tasks related to physical and social sciences have increased by a third (from 6% to 8%). Meanwhile, the relative frequency of traditional business tasks has declined: management-related tasks have decreased from 5% to 3% of all conversations, and tasks related to business and financial operations have halved, from 6% to 3%. (Of course, in absolute terms, the number of conversations in each category has still increased significantly.)

The overall trend shows that as a country's per capita GDP increases,ClaudeApplications are migrating from computer/mathematical tasks to more diverse scenarios such as education, arts and design, administrative support, and scientific research. Software development remains the most frequently used application in all countries, and a similar trend is observed across US states, although our sample size limits our ability to explore in more detail how the task mix changes with adoption rates.

Interaction mode

As weBeforeAs mentioned,we willAIApplications are divided intoautomaticchange(AIMinimize direct output from human intervention andEnhance(Human-machine collaboration), we will furtherautomaticThe enhancement is further subdivided into imperative dialogue and feedback loop interaction. Imperative dialogue involves very little human-computer interaction, while in feedback loop tasks, humans provide real-world results to the model. We also further subdivide enhancement into learning (requesting information or explanation) and task iteration (and...). Claude Collaboration) and verification (requesting feedback).

Since December 2024, the proportion of command-based dialogues has surged from 27% to 39%, while other modes (especially learning, task iteration, and feedback loops) have slightly decreased, driving...automaticThe total amount of chemical oxidation (49.1%) surpassed the enhancement (47%) for the first time. This indicates...AIjustfastGaining user trust and increasing autonomy to undertake more complex tasks.

The main reason is likely the upgrade in model capabilities (when we first collected economic index data in December 2024).Claude ofup to dateThe version is Sonnet 3.6. As the model becomes better at predicting user needs and generating high-quality results, users may be more willing to trust the model's output on their first try.

The anomaly is that countries with high usage rates tend to favor the enhancement model, while countries with low usage rates prefer...automaticUnder controlled task combinations, for every 1% increase in average usage per user,automaticThe adoption rate decreased by approximately 3%. The reasons are unclear, but may be related to cultural differences or economic factors among early users in different countries.

Use andClaudeUsing the same privacy-preserving approach for .ai conversations, we sampled interactions from a subset of Anthropic's first-party API clients in a groundbreaking analysis. These API clients are typically enterprises and developers who use... Claude Methods and passagesClaude.ai access Claude They are very different: they pay by tokens, not by a fixed monthly subscription, and can make requests through their own programs.

These customers Claude Its use is particularly focused on coding and management tasks: in our example, 44% of API traffic is mapped to computer or mathematical tasks, whileClaudeOn .ai, this figure is 36%. (In reality, approximately 5% of API traffic is dedicated to development and evaluation.)artificialintelligentThe proportion of conversations related to education and careers (in the API is 4%).Claude.aiThe percentage of conversations related to arts and entertainment (12%) and the percentage of conversations related to arts and entertainment (5% in the API) are also mentioned.ClaudeThe relatively small percentage (8% on .ai) offset the difference.

Finally, considering the payment methods for API usage, we explored whether differences in task costs (caused by differences in the number of tokens consumed) affect which tasks enterprises choose to "purchase." We found a positive correlation between task cost and usage frequency: task categories with higher costs tend to be used more frequently, as shown in the figure below. This indicates that the model's basic functionality and the economic value it generates are more important to enterprises than the cost of completing the task itself.

Economic indices are designed to empirically assessAIIn the initial stages of its impact on employment and the economy, we found that...AIThe applications exhibit a significant imbalance: users in high-income countries tend to use...ClaudeMore emphasis on collaboration thanautomaticIts application extends beyond programming. In the US, its usage at the state level is closely linked to local leading industries (from technology to tourism), and businesses are more willing to assign it to users than consumers.ClaudeSovereignty.

Of particular note is the past nine monthsClaude.ai command typeautomaticThe proportion of interactive dialogues has increased dramatically. Users are still exploring and...AIThe boundaries of collaboration—we are still collectively defining the boundaries.AITrust thresholds and responsibility allocation for tools. However, current trends indicate that humans are becoming increasingly adapted...AIAgent work. We will continue to track this evolution and observe whether user behavior tends to stabilize as the model's capabilities improve.

Welcome to the Anthropic Economic Index website to explore interactive data visualizations (covering country, state, and occupational dimensions). We will continuously update the data for public tracking.AIThe path of its impact on the economy and employment.

Open Data

As with previous reports, this issue also includes the complete dataset, which includes geographic data, task-level usage patterns, and task-specific data.automaticA breakdown of enhancements/features and an overview of API usage. Data is available...Economic Index Websitedownload.

Report official website

Report official websitehttps://www.anthropic.com/research/economic-index-geography
Interactive websitehttps://www.anthropic.com/economic-index

All Categories