Anthropic Economic Index report: Learning curves

The Anthropic Economic Index utilizes a privacy-preserving data analysis system to monitor the usage of Claude across various sectors of the economy. This initiative aims to understand the economic impacts of artificial intelligence (AI) early on, providing researchers and policymakers with sufficient time to prepare for changes. This report focuses on Claude’s usage during February 2026, building upon the economic primitives framework introduced in the previous report from November 2025.

Overview of Claude Usage in February 2026

The analysis covers the period from February 5 to February 12, coinciding with the releases of Claude Opus 4.5 and 4.6. The findings reveal several key trends:

• The rate of augmentation, or collaborative interaction where AI complements user abilities, slightly increased in both Claude.ai and API traffic.
• Usage on Claude.ai diversified, with the top 10 tasks accounting for a smaller share of overall usage compared to November 2025.
• The average conversation in Claude.ai involved slightly lower-wage tasks than in previous reports.

Changes Since the Last Report

In the first chapter of this report, we revisit findings from the previous Economic Index report published in January 2026. Key observations include:

• Diversification of Use Cases: Coding tasks are transitioning from augmentative usage in Claude.ai to more automated workflows in API traffic. The top 10 tasks in Claude.ai made up 19% of all traffic in February, down from 24% in November 2025.
• Broadened Adoption: The average economic value of work performed on Claude has decreased slightly, attributed to an increase in personal queries related to sports, product comparisons, and home maintenance.
• Persistent Inequality: Global usage remains concentrated, with the top 20 countries accounting for 48% of all per-capita usage, an increase from 45%. However, within the United States, usage per capita continued to converge, with the share of usage from the top 10 states decreasing from 40% to 38%.

Learning Curves in Claude Adoption

A central finding of the Economic Index is that the early adoption of Claude is uneven. It is more intensely used in high-income countries, particularly in areas with a higher concentration of knowledge workers, and for a limited range of specialized tasks and occupations. An important question arises: how might this inequality in adoption affect the distribution of AI benefits?

If effective AI usage requires complementary skills that can be developed through experience, the advantages of early adoption may become self-reinforcing. In the second chapter, we explore how users shape the value they derive from Claude by matching model capabilities to tasks and how their usage patterns evolve with experience.

Model Selection and Task Matching

Our analysis shows that users tend to select the most advanced model class, Opus, for tasks that typically command higher wages in the labor market. For instance:

• Among paying Claude.ai users, Opus is utilized 4 percentage points more than average for coding tasks.
• Conversely, Opus is used 7 percentage points less than average for tutoring-related tasks.

This model-switching behavior is even more pronounced among API users.

Experience and Success Rates

Our findings indicate that seasoned Claude users tend to engage with the platform for higher-value educational tasks and less frequently for personal inquiries. For example:

• Users with six months or more of experience have 10% fewer personal conversations.
• They also show a 6% higher education level in their inputs.
• Most notably, this group has a 10% higher success rate in their interactions, which cannot be solely attributed to task selection, country of origin, or other factors.

This trend may reflect the sophistication of early adopters or suggest a phenomenon of learning-by-doing, where users improve their skills through ongoing experience with Claude.

Diversification of Use Cases in Claude.ai

In examining the types of tasks Claude is asked to perform, we utilized our privacy-preserving system to analyze behavior at an aggregated level without revealing individual conversation content. We sampled 1 million conversations from both Claude.ai and our first-party API.

Coding tasks remain the most prevalent on our platforms, accounting for 35% of conversations on Claude.ai. However, between November 2025 and February 2026, we observed a decrease in task concentration:

• The top 10 most common O*NET tasks declined from 24% to 19% of conversations.

This decline is partly due to coding tasks migrating to our first-party API, where Claude Code has gained significant traction. Claude Code’s architecture allows for coding work to be divided into smaller API calls, thus spreading the workload across various task categories.

Conclusion

In summary, the Anthropic Economic Index report highlights the evolving landscape of AI usage through Claude. As adoption diversifies and users develop their skills, the implications for the labor market and economic value are significant. The findings underscore the importance of understanding how AI can be harnessed effectively to maximize its benefits across different sectors.

Note: The insights presented in this report are based on data collected during February 2026 and may evolve as further research is conducted.