All occupations in each band added together, May 2022 to May 2025.
Same, grouped by each occupation's share of Claude conversations.
Eloundou, Manning, Mishkin and Rock had both human annotators and GPT-4 label each of about 19,000 O*NET task statements. E1, direct: a model like ChatGPT alone could cut the time to do the task at equal quality by at least half. E2, with tools: the model alone could not, but software built on top of it could. E0: neither. The occupation score used here is their β measure from the GPT-4 labels, which counts E1 tasks fully and E2 tasks at half. It is the share of an occupation's tasks that are exposed, from 0 to 1. Where one BLS occupation covers several O*NET occupations, the score is the plain average of theirs. Each task also shows what the human annotators said, because the two often disagree.
The Anthropic Economic Index classifies a privacy-preserving sample of Claude conversations against the same O*NET tasks. The figure is each task's or occupation's share of all sampled conversations in , as published, rounded to two decimals. It is a share of conversations, not of workers. About of Claude.ai conversations that month were classified as work; the rest were personal or coursework, and they are counted against occupations too. "Automation" is Claude doing the task with little back-and-forth (directive or feedback-loop conversations); "augmentation" is the person iterating, learning or checking. A blank means Anthropic did not publish the cell because too few conversations fell in it, which is not the same as zero.
Each conversation is matched to one task statement, so occupations that own a very generic statement collect a lot of usage. Librarians rank near the top because "search standard reference materials, including online sources and the Internet, to answer patrons' reference questions" is what a lot of people ask Claude to do, not because librarians are heavy users. "Computer occupations, all other" tops the chart for the same reason: one of its O*NET occupations, document management specialists, owns "search electronic sources, such as databases or repositories, or manual sources for information."
Employment and median annual wage are the BLS Occupational Employment and Wage Statistics for May 2022 (the last estimates gathered entirely before ChatGPT's release that November) and May 2025. Each OEWS estimate pools six survey panels over three years, so the two estimates here draw on separate panels, but BLS does not recommend OEWS for time series: methods and occupational coding change, and older panels' wages are aged forward with a model. Wages are nominal; prices rose over the same three years. A change in any one occupation could come from interest rates, the end of pandemic hiring, offshoring or a dozen other things. Nothing on this page shows that LLMs caused it.
47 "All other" occupations (2.0 million jobs in May 2025) have no O*NET tasks and so no
exposure score. Military occupations and fishing and hunting workers are not in OEWS. Twelve
OEWS codes that combine several detailed occupations are matched to all the O*NET occupations
under them. Three task statements that O*NET retired in release 31.0 but that Claude
conversations are still classified into are kept and marked "retired". The full list is in the
app's scripts/dropped.txt.
full_labelset.tsv) and occupation scores (occ_level.csv) from
github.com/openai/GPTs-are-GPTs.aei_claude_ai_2026-06-26.csv and aei_1p_api_2026-06-26.csv,
huggingface.co/datasets/Anthropic/EconomicIndex (CC-BY).oesm22nat.zip, oesm25nat.zip),
bls.gov/oes/tables.htm; on comparing years, see
BLS, "Can OEWS data be used to look at changes in employment and wages over time?",
bls.gov/oes/oes_ques.htm.