Anthropic Economic Index: AI’s impact on software development

Jobs that involve computer programming are a small sector of the modern economy, but an influential one. The past couple of years have seen them changed dramatically by the introduction of AI systems that can assist with—and automate—significant amounts of coding work.

In our previous Economic Index research , we found very disproportionate use of Claude by US workers in computer-related occupations: that is, there were many more conversations with Claude about computer-related tasks than one would predict from the number of people working in relevant jobs. It’s the same in the educational context : Computer Science degrees—which involve large amounts of coding—show highly disproportionate AI use.

To understand these changes in more detail, we conducted an analysis of 500,000 coding-related interactions across Claude.ai (the “default” way that most people interact with Claude) and Claude Code (our new specialist coding “agent” that can independently accomplish chains of complex tasks using a variety of digital tools).

Conversations that related to more generic uses, such as “Software Architecture & Code Design” and “Debug and Performance Optimization” were also highly represented in both Claude.ai and Claude Code.

We also analyzed which groups of developers might be using Claude. We used our analysis system to identify the type of project (e.g. a personal project vs. a project done for a startup) that best described users’ coding-related interactions. Because we don’t know the real-world context in which Claude’s responses were being used, these analyses rely on uncertain inferences from incomplete data. We therefore treat these findings as more preliminary than the ones described above.

Our analysis is grounded in real-world AI use—how developers are actually using Claude in their workflows. Although this approach gives our findings practical relevance, it also brings inherent limitations. These include:

Compared to use cases that don’t involve software, software development is more automative. A significant increase in Feedback Loops (+18.3%) drives this and, notably, offsets a clear decrease in Directive behaviors (-11.2%). In other words, AI-assisted coding currently requires a lot of human reviewing and iteration relative to non-coding tasks, even when Claude does the bulk of the work.

We analyzed the 500,000 total Claude interactions (split between Claude Code and Claude.ai 1 ) using our privacy-preserving analysis tool , which distills user conversations into higher-level, anonymized insights. Here, we used it to identify the topic of the conversation (e.g. “UI/UX component development”), or—as we’ll explain below—to categorize a conversation as focusing on “augmentation” versus “automation”.

In our previous Economic Index reports, we separated out “automation,” where AI directly performs tasks, from “augmentation,” where AI collaborates with a user to perform a task. Here, we found that Claude Code showed dramatically higher automation rates—79% of conversations involved some form of automation, compared to 49% on Claude.ai.

We also split automation and augmentation into several subtypes (as discussed in our previous work ). “Feedback Loop” patterns, where Claude completes tasks autonomously but with help of human validation (for example, where the user sends any errors back to Claude), were nearly twice as common on Claude Code (35.8% of interactions) as Claude.ai (21.3%). “Directive” conversations, where Claude completed a task with minimal user interaction, were also higher on Claude Code (43.8%, versus 27.5% on Claude.ai). All the patterns of augmentation—including “Learning,” where the user acquires knowledge from the AI model—were substantially lower on Claude Code than on Claude.ai.