The first Anthropic Economic Index report (February 2025) measured which occupational tasks people use Claude for, using conversations from December 2024 and January 2025, and promised to repeat the measurement as models changed. This short post, the second report, does so after the launch of Claude 3.7 Sonnet, a model Anthropic described as having strengths in agentic coding and a new "extended thinking" mode. It asks three questions: did the mix of tasks change, what is extended thinking used for, and how do automation and augmentation vary by task and occupation. It also releases a new dataset built from the conversations themselves rather than from a predefined task list. All numbers below are the post's; there is no accompanying paper.
What the report does
The authors reran the first report's pipeline on 1 million anonymized Claude.ai Free and Pro conversations from the 11 days after Claude 3.7 Sonnet's launch, most of them with that model, since it was the default on Claude.ai and the mobile app (section "What's changed since the launch of Claude 3.7 Sonnet?"). As before, the privacy-preserving tool Clio maps each conversation to a task in the U.S. Department of Labor's O*NET database, which the post describes as having 17,000 tasks, and those tasks to occupations and occupational categories.
Two methods changed, as the post documents in its final section, "Additional methodological details":
- No occupational screen. The first report kept only conversations that a classifier judged relevant to an occupational task. This report instead removes only conversations flagged by safety classifiers. The authors say the two approaches give similar results, and the change keeps more data for the new bottom-up dataset.
- Newer classifier model. Claude 3.7 Sonnet replaces Claude 3.5 Sonnet wherever the pipeline previously used Claude 3.5 Sonnet, because it was more accurate on the internal benchmarks used to assess Clio. No accuracy figures are given.
What it finds
A modest shift toward coding, education and science
The share of usage rose modestly for coding, education and the sciences (the post's opening summary also lists healthcare applications). Coding growth was expected from Claude 3.7 Sonnet's better coding benchmark scores; the authors offer three possible explanations for the other increases: continued diffusion of AI, new applications of coding in those fields, or unexpected capability gains. The appendix chart shows Computer and Mathematical occupations with the largest absolute increase, given in the chart caption as "+3%". The post stresses that these are shares of usage, not absolute amounts.
The depth-of-use curve, which plots the fraction of occupations against the fraction of their tasks with any AI use, is very similar to the first report's; the chart caption gives the example that about 40% of occupations show use in at least 20% of their tasks. If anything, the area under the curve is slightly smaller, which the authors suggest may reflect usage concentrating toward coding (appendix, "Task Curve").
Who uses extended thinking
Extended thinking is used most in technical and creative problem-solving. Among occupations with at least 0.5% of the data, tasks associated with computer and information research scientists lead, with almost 10% of conversations using extended thinking, followed by software developers at around 8%, multimedia artists at about 7% and video game designers at about 6% (section "How are people using extended thinking mode?"). The report releases the fraction of thinking-mode use for each O*NET task.
Automation and augmentation by task and occupation
Using the first report's five collaboration patterns (Directive and Feedback Loop count as automation; Task Iteration, Learning and Validation as augmentation), the overall balance is "essentially unchanged," with augmentation at 57%. Within it, Learning conversations, where the user asks for information or explanation, rose from about 23% to about 28% (section "How does augmentation vs. automation vary by task and occupation?").
By occupational category, Community and Social Service tasks, which include education and guidance counseling, approach 75% augmentation; Production and Computer and Mathematical tasks are closer to 50-50. No category is dominated by automation.
By occupation, among those with at least 0.5% of the data:
- Copywriters and editors show the most Task Iteration, where the user and model write together; the chart caption gives about 58% for copywriters.
- Translators and interpreters show among the most Directive use, with documents translated with minimal human involvement.
- Librarians show the highest share of Learning, about 56% according to the chart caption.
These task- and occupation-level breakdowns were released in response to requests through the first report's researcher input form. The post cautions that O*NET descriptions may not fit actual use: conversations classed under "fine artists, including painters, sculptors, and illustrators" probably involve digital art far more than painting or sculpture.
A bottom-up taxonomy
Because ONET may not cover everything a general-purpose model is used for, the report releases a dataset built bottom-up by Clio from the same conversations: 630 clusters in a three-level hierarchy, each with a description, prevalence and automation and augmentation breakdown (section "A bottom-up taxonomy of usage on Claude.ai"). Using the same conversations makes the top-down ONET results and the bottom-up clusters directly comparable. The post says the categories range from "Help resolve household plumbing, water, and maintenance issues" to "Provide guidance on battery technologies and charging systems," and highlights others such as "Help with time zone handling in code and databases." Analysis of the dataset is left to future work.
The prompts and classifiers
The report prints no prompts and changes the classification setup in the two ways described above. Because the automation and augmentation figures come from a model reading conversations against written definitions of each pattern, a change of classifier model is a change in the measuring instrument as well as in the data. The post reports only that the newer model was more accurate on internal benchmarks, not how much the change alone moves any of the reported shares.
Limitations and open questions
The post states few limitations of its own beyond the screening change and the fit of O*NET descriptions; the first three points below follow from its design rather than from the authors.
- Two changes at once. Both the data period and the pipeline (screening rule and classifier model) changed between reports, so differences such as the rise in Learning cannot be attributed to the new model or to user behavior from the post alone. The authors say the screening change gives similar results but do not quantify it.
- Shares, not volumes. All comparisons are of shares of usage; growth in one category lowers the others' shares.
- Short window. Eleven days after a launch may reflect novelty in how people try a new model.
- Sample. As in the first report, only Claude.ai Free and Pro conversations are included, and O*NET's categories may not describe what people actually do.
- Later data. The September 2025 report, Uneven geographic and enterprise AI adoption, compares its data with both this report's sample and the first, and states that directive automation rose from 27% to 39%.
Research questions
- How much of the change between reports comes from the pipeline? Rerunning the first report's December 2024 sample through this report's pipeline (no occupational screen, Claude 3.7 Sonnet as classifier) would separate measurement changes from changes in use.
- Why did Learning conversations grow? The share rose from about 23% to 28%. Comparing Learning shares by model and by whether extended thinking was on would show whether the new model or new users drove it.
- Does extended thinking change the collaboration pattern? The report gives thinking-mode fractions by task. Crossing them with collaboration patterns would show whether users who switch on extended thinking delegate more or iterate more.
- How much usage falls outside O*NET? The bottom-up taxonomy enables this. Matching each of the 630 clusters to its nearest O*NET task and measuring the share of usage in poorly matched clusters would quantify the gap the report names.
- Are occupation-level patterns stable over time? Tracking Task Iteration for copywriters and Directive use for translators in later reports would show whether these are stable features of the work or snapshots.
How it connects
- First report. The Anthropic Economic Index (February 2025) defines the pipeline and collaboration patterns this report reuses.
- Later reports. Uneven geographic and enterprise AI adoption (September 2025) uses this report's February to March 2025 data as one of its comparison points.
- Infrastructure. Clio (December 2024) produces both the O*NET mapping and the bottom-up clusters.
References
- Anthropic. (2025, March 27). Anthropic Economic Index: Insights from Claude 3.7 Sonnet. https://www.anthropic.com/news/anthropic-economic-index-insights-from-claude-sonnet-3-7
- Dataset: https://huggingface.co/datasets/Anthropic/EconomicIndex