Anthropic reveals how people actually use Claude 3.7 Sonnet

Everyone has an opinion on whether AI is coming for our jobs or just acting as an overqualified intern. Anthropic decided to skip the punditry and look at the receipts.

The AI lab just released its second Anthropic Economic Index report, breaking down one million anonymized Free and Pro conversations on Claude.ai in the eleven days following the launch of Claude 3.7 Sonnet. Using their privacy-preserving Clio tool, they mapped real-world user prompts against 17,000 tasks tracked by the U.S. Department of Labor's O*NET database.

Extended thinking goes to the nerds, while humans keep the wheel

When Anthropic rolled out Claude 3.7 Sonnet, the headline feature was its hybrid extended thinking mode, allowing the model to deliberate before generating an answer. Unsurprisingly, people are not turning on deep thinking to write thank-you notes.

Extended thinking is heavily concentrated in technical problem-solving. It was triggered in nearly 10% of tasks tied to computer and information research scientists, followed by software developers at roughly 8%. Digital creative roles like multimedia animators at around 7% and video game designers at 6% also leaned on the toggle frequently.

Meanwhile, the narrative of total workforce displacement took another hit. The breakdown between augmentation and automation held dead flat at 57% in favor of augmentation. In fact, learning interactions—where users ask Claude to explain a topic rather than just spit out a finished product—rose from 23% to 28%.

Across all occupational categories analyzed, not a single one was dominated by full automation. Community and social service tasks skewed nearly 75% toward augmentation. Even technical fields like computer and math occupations hovered around a 50-50 split.

The task-level dynamics are telling. Copywriters and editors showed the highest back-and-forth iteration, essentially co-writing drafts with Claude. Translators, on the other hand, showed the highest directive behavior, handing off the text and letting the model finish the job solo.

630 bottom-up ways people actually use Claude

Traditional labor classifications often miss how messy and flexible real AI workflows actually are. To fix this blind spot, Anthropic released an entirely new, bottom-up taxonomy covering 630 granular clusters.

Instead of broad labels, the dataset captures real edge cases. Prompts ranged from debugging timezone issues in databases and advising on battery systems, all the way to troubleshooting household plumbing and picking typography. Both the taxonomy and task-level thinking datasets are now freely available on Hugging Face.

Why it matters

Anthropic is giving us our cleanest look yet at how advanced reasoning models actually fit into the daily grind. The data clearly shows that users treat models like Claude 3.7 Sonnet as thinking partners rather than hands-off replacements. If you are a knowledge worker, the takeaway is reassuring: the winning strategy right now is iterative co-creation, not looking for a magical one-click button.

Publishing this data is also a brilliant strategic move for Anthropic. It anchors the frontier model race in tangible economic utility rather than speculative AGI hype, while quietly highlighting Sonnet's coding and reasoning chops. The caveat is that this study only tracks conversational UI traffic. It completely misses headless API pipelines, where real automated agent workflows are far more likely to be hiding.