Anthropic Economic Index: Insights from Claude 3.7 Sonnet

Anthropic has released its second research report under the Anthropic Economic Index, analyzing usage patterns of Claude 3.7 Sonnet across 1 million anonymized conversations. The findings indicate a modest increase in AI adoption within coding, education, and the sciences, while the balance between augmentative and automative usage remains stable.

Shift in Occupational Usage Patterns

Since the launch of Claude 3.7 Sonnet, there has been a measurable increase in the share of usage within several occupational categories. Computer and mathematical occupations saw the largest absolute increase (+3%), with notable percentage increases also appearing in education and the sciences.

Anthropic attributes the rise in coding usage to the model's improved benchmark scores, while the growth in other sectors may stem from the broader diffusion of AI, the application of coding to new domains, or general capability improvements in the model.

Usage of Extended Thinking Mode

Claude 3.7 Sonnet's "extended thinking" mode is predominantly utilized for technical and creative problem-solving. According to the data, the highest usage rates for this mode are found among:

  • Computer and Information Research Scientists: ~10% usage
  • Software Developers: ~8% usage
  • Multimedia Artists: ~7% usage
  • Video Game Designers: ~6% usage

Augmentation vs. Automation

Augmentation continues to be the primary way users interact with AI, comprising 57% of total usage. While the overall balance between augmentation and automation is unchanged, the share of "learning" interactions—where users seek information or explanations—rose from approximately 23% to 28%.

Interaction Patterns by Occupation

Different occupations exhibit distinct interaction styles based on the nature of their tasks:

  • High Augmentation: Community and Social Service tasks (including education and guidance counseling) show the highest augmentation rates, approaching 75%.
  • Task Iteration: Copywriters and editors show the highest proportion of task iteration, where humans and the model co-write content (leading at ~58%).
  • Directive Behavior: Translators and interpreters exhibit some of the highest levels of directive behavior, where the model completes tasks with minimal human involvement.
  • Learning: Librarians show the highest proportion of learning interactions at approximately 56%.

Anthropic notes that O*NET descriptions may not perfectly align with actual usage; for example, usage attributed to "fine artists" likely refers to digital art rather than traditional sculpture or painting.

Bottom-Up Usage Taxonomy

To capture use cases that fall outside the US Department of Labor's O*NET database, Anthropic introduced a bottom-up taxonomy consisting of 630 granular clusters. This dataset identifies specific, niche applications of Claude.ai, including:

  • Guidance on battery technologies and charging systems.
  • Creation of physics-based simulations with interactive visualization.
  • Help with time zone handling in code and databases.
  • Support for water management systems and infrastructure projects.
  • Font selection, implementation, and troubleshooting.

Methodology and Data Availability

Anthropic used its privacy-preserving analysis tool, Clio, to map conversations to O*NET tasks. For this second report, the lab replaced Claude 3.5 Sonnet with Claude 3.7 Sonnet for classifications, which improved accuracy according to internal benchmarks. The analysis filtered out conversations flagged by safety classifiers but did not filter based on occupational relevance to ensure a more comprehensive bottom-up taxonomy.

The datasets for these analyses, including the O*NET task-to-thinking-mode mapping and the bottom-up taxonomy, are freely available on Hugging Face.

Sources