Anthropic Economic Index Insights From Claude Sonnet 3 7

Anthropic Economic Index: Insights from Claude 3.7 Sonnet \ Anthropic Societal Impacts Economic Research Anthropic Economic Index: Insights from Claude 3.7 Sonnet Mar 27, 2025

Last month, we launched the Anthropic Economic Index —a new initiative where we’re regularly releasing data and research aimed at understanding AI's effects on labor markets and the economy over time. Today, we’re releasing our second research report from the Index, covering usage data on Claude.ai following the launch of Claude 3.7 Sonnet —our newest and most capable model with strengths in agentic coding and a new “extended thinking” mode. Briefly, our latest results are the following: Since the launch of Claude 3.7 Sonnet, we’ve observed a rise in the share of usage for coding, as well as educational, science, and healthcare applications; People use Claude 3.7 Sonnet’s new “extended thinking” mode predominantly for technical tasks, including those associated with occupations like computer science researchers, software developers, multimedia animators, and video game designers; We're releasing data on augmentation / automation breakdowns on a task- and occupation-level. For example, tasks associated with copywriters and editors show the highest amount of task iteration , where the human and model co-write something together. By contrast, tasks associated with translators and interpreters show among the highest amounts of directive behavior—where the model completes the task with minimal human involvement.

In addition, we’re releasing a first-of-its-kind bottom-up taxonomy of usage on Claude.ai. This new dataset covers 630 granular categories ranging from “Help resolve household plumbing, water, and maintenance issues” to “Provide guidance on battery technologies and charging systems.” We hope this bottom-up taxonomy will be useful for researchers, and reveal use-cases that might be missed by top-down approaches which map usage onto a list of predefined tasks. The datasets for these analyses are freely available to download . Read on for more details on our findings. What’s changed since the launch of Claude 3.7 Sonnet? Last month, we introduced Claude 3.7 Sonnet, our most capable model yet with an “extended thinking mode”. We reran our previous analysis on data from the 11 days following the launch, covering 1 million anonymized Claude.ai Free and Pro conversations. The vast majority of the data we analyzed was from Claude 3.7 Sonnet, as it is set as the default on Claude.ai and our mobile app. As a reminder, our privacy-preserving analysis tool, Clio , maps each conversation to one of 17,000 tasks in the U.S. Department of Labor’s O*NET database. We then look at the overall patterns in the occupations and high-level occupational categories associated with those tasks. When looking at the breakdown of these 1 million conversations, we see that the proportion of usage in several occupational categories has increased modestly, including coding, education and the sciences. While this increase in coding usage was expected due to Claude 3.7 Sonnet’s improved scores on coding benchmarks, the increase in these other categories could reflect either ongoing diffusion of AI throughout the economy, novel applications of coding to those domains, or unexpected capability improvements in the model. In the two months since our original data sample, we’ve seen an increase in the share of usage for coding, education, and the sciences. Graph shows share of Claude.ai Free and Pro traffic across top-level occupational categories in O*NET. Grey shows the distribution from our first report covering data from Dec ‘25 - Jan ‘25. Colored bars show an increase (green) and decrease (blue) in the share of usage for our new data from Feb ‘25 - March ‘25. Note that the graph shows the share of usage rather than absolute usage. See Appendix for chart showing change across the full list of occupational categories.

How are people using extended thinking mode? Claude 3.7 Sonnet features a new “extended thinking” mode which, when activated by the user, enables the model to think for longer when answering more complex questions. Our analysis reveals that Claude 3.7 Sonnet's extended thinking mode is predominantly used in technical and creative problem-solving contexts. Tasks associated with computer and information research scientists lead with almost 10% using extended thinking, followed by software developers at around 8%. Tasks associated with digital creative roles like multimedia artists (~7%) and video game designers (~6%) also show substantial usage. While these early usage patterns reveal insights about when people choose to use extended thinking mode, many important questions remain about this new model capability. To enable further research in this space, we’re releasing a new dataset that maps each O*NET task to its associated thinking mode fraction. This dataset is available on our Hugging Face page . What tasks see the highest associated usage of extended thinking mode? Graph shows the O*NET occupations with highest usage of thinking mode in their associated tasks. Occupations shown are limited to those with at least 0.5% representation in the data. How does augmentation vs. automation vary by task and occupation? In our last report, we analyzed how AI usage varied between augmentative uses, like learning or iterating on an output, and automative uses, like asking the model to directly complete a task or debug errors. Our analysis shows the balance of augmentation and automation is essentially unchanged in our new data, with augmentation still comprising 57% of usage. However, we did see some change in types of automation and augmentation uses—for example, we saw learning interactions, where the user asks Claude for information or explanation about different topics, rise from ~23% to ~28%.

The balance of augmentation and automation has stayed relatively constant in the two months between our data samples (V1 and V2), though the share of Learning conversations has grown appreciably.

We received a number of requests via our researcher input form to release automation and augmentation data at the level of tasks and occupations. We do just that in this report, providing this data on our Hugging Face page . When splitting the data by high-level occupational categories, we find some categories are highly augmentative; for example, Community and Social Service tasks, which includes education and guidance…