What Changed After Claude 3.7 Sonnet Launched?
Anthropic’s second Economic Index report, released on March 27, 2025, analyzed 1 million anonymized Claude.ai Free and Pro conversations from the 11 days following the launch of Claude 3.7 Sonnet. The goal was to understand how people actually use a more capable model, not just how it performs on benchmarks. The report builds on the first Economic Index, which covered data from December 2025 to January 2025, and uses the same privacy-preserving tool, Clio, to map conversations to tasks in the U.S. Department of Labor’s O*NET database.
Compared to the earlier sample, the share of usage in coding, education, and science increased modestly. Anthropic notes that the coding increase was expected because Claude 3.7 Sonnet scores higher on coding benchmarks. The rise in education and science could reflect broader AI diffusion, novel applications of coding to those domains, or unexpected capability improvements. Importantly, the report shows shares of usage, not absolute numbers. A category’s share can drop even if its absolute usage grows, simply because other categories grew faster. For product builders, this distinction matters: a decline in one area doesn’t necessarily mean users abandoned it.
The report also highlights that computer and mathematical occupations saw the largest absolute increase (+3%), while education and the sciences showed notable percentage increases. This suggests that AI is moving beyond pure software development into adjacent fields, which could signal new opportunities for tools that bridge coding with domain-specific workflows.
How People Use Extended Thinking Mode
Claude 3.7 Sonnet introduced an “extended thinking” mode that lets the model reason longer on complex questions. The report reveals that this mode is predominantly used for technical and creative problem-solving. Tasks associated with computer and information research scientists lead with almost 10% of usage involving extended thinking, followed by software developers at around 8%. Digital creative roles also show substantial usage: multimedia artists at ~7% and video game designers at ~6%.
These numbers suggest that users self-select extended thinking when they face tasks that require deeper reasoning, such as algorithm design, complex debugging, or creative simulation. For product builders, this is a clear signal: if your tool can detect when a user is tackling a complex task, you might automatically suggest or enable a similar “deep thinking” mode. The data also implies that users are willing to wait longer for better answers on hard problems, which could inform UX decisions around latency and progress indicators.
Anthropic released a new dataset that maps each O*NET task to its associated thinking mode fraction, available on Hugging Face. This allows researchers and product teams to analyze which specific tasks drive extended thinking usage, potentially revealing niche opportunities for specialized AI features.
Augmentation vs. Automation: The Balance Holds
The report distinguishes between augmentative uses (learning, iterating on outputs) and automation uses (directly completing tasks, debugging). The overall balance remains unchanged: augmentation still comprises 57% of usage. However, the types of augmentation shifted. Learning interactions, where users ask Claude for information or explanations, rose from ~23% to ~28% of all usage. This growth in learning suggests that people are increasingly using AI as an educational tool, not just a task executor.
Across occupational categories, the augmentation-automation split varies widely. Community and Social Service tasks, which include education and counseling, approach 75% augmentation. In contrast, production or computer and mathematical occupations skew closer to 50-50%. Notably, no occupational category shows automation dominating. This is a key insight for product strategy: most users still want to collaborate with AI rather than hand off entire tasks.
At the task level, copywriters and editors show the highest amount of “task iteration,” where human and model co-write content together. Translators and interpreters show among the highest “directive” behavior, where the model completes translations with minimal human involvement. These examples illustrate that the same model can support different collaboration styles depending on the domain. Product builders should consider designing interfaces that support both iterative co-creation and more autonomous execution, depending on the user’s goal.
A Bottom-Up Taxonomy of Real Usage
The report also introduces a first-of-its-kind bottom-up taxonomy of Claude.ai usage, created with Clio from the same anonymized conversations. This dataset contains 630 granular clusters, organized into three levels of hierarchy, with descriptions, prevalence metrics, and automation/augmentation breakdowns. Unlike top-down approaches that map usage onto predefined ONET tasks, this bottom-up approach captures use cases that ONET might miss.
Some interesting clusters include: helping with water management systems and infrastructure projects, creating physics-based simulations with interactive visualization, assisting with font selection and troubleshooting, creating or improving job application materials, providing guidance on battery technologies and charging systems, and handling time zone issues in code and databases. These examples show the diversity of real-world applications, many of which are not obvious from traditional job descriptions.
For product builders, this taxonomy is a goldmine. It offers a data-driven view of what people actually do with AI, which can inspire new features or identify underserved niches. The dataset is freely available on Hugging Face, so teams can analyze it directly to inform their product roadmaps.
Limitations and Practical Takeaways
The report has several limitations. The sample is limited to Claude.ai Free and Pro conversations over just 11 days, so it may not generalize to the entire labor market or to other AI products. Anthropic also notes that O*NET descriptions may not perfectly represent Claude’s actual usage. For example, the occupation “fine artists, including painters, sculptors, and illustrators” likely involves more digital art than traditional painting or sculpture on Claude.ai. Additionally, the report uses shares of usage, not absolute numbers, which can be misleading if not interpreted carefully.
For product builders, the key takeaways are:
- Extended thinking is a feature users want for hard problems. Consider adding a similar mode that activates automatically for complex tasks, or at least make it easy to enable.
- Augmentation dominates automation. Most users want to collaborate with AI, not replace themselves. Design workflows that support iteration, learning, and feedback loops.
- Learning is growing. The rise in learning interactions suggests that educational features, explanations, and tutorials can be valuable additions to AI tools.
- Use the released datasets. The bottom-up taxonomy and thinking-mode data are publicly available. Analyze them to discover real use cases and align your product with actual demand.
In conclusion, Anthropic’s second Economic Index report provides a rare, data-rich view of how people use a frontier AI model. While it’s not a perfect mirror of the entire economy, it offers actionable insights for anyone building AI products. By understanding what tasks drive extended thinking, where augmentation vs. automation dominates, and what real use cases emerge bottom-up, product teams can make more informed decisions about features, UX, and positioning. The datasets are open, so the next step is to dig in and see what patterns you can find for your own product.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.