Anthropic's latest AI Economic Index report, based on a survey of 9,700 users and telemetry data, reveals how AI is affecting work. The higher the job's value, the more compute it consumes: high-paying roles use 2.07 times more compute than low-paying ones. For example, a marketing manager writing a proposal uses 2.5 times the compute of an editor revising an article. But there are exceptions: a high-earning pharmacist uses only one-twentieth the compute of a statistical assistant.
Over a third of respondents expect AI to take over most of their work within a year. Surprisingly, the people who delegate the most to AI are the most optimistic about their future income and job prospects, and aren't worried about skill decay. Autonomy varies by tool: on the web, users go back and forth with AI an average of 13 rounds when writing; with Claude Code (a terminal tool), they typically give a single command and let AI generate the output in one go.
In autonomy scoring, users of Claude Code are more willing to let AI make decisions on its own — their score is 0.37 points higher than web users (0.26 higher for the same model). The exception is working with spreadsheets: web users often do financial modeling that requires planning, while Claude Code users mostly do mechanical data extraction. So for spreadsheet tasks, web AI actually gets a 0.35-point higher autonomy score. AI's final responses are generally deeper than the user's initial query; in design and game development tasks, the education level required to understand AI's reply is almost two years higher than the user's question.
Telemetry also reveals overtime patterns: during nights and weekends, high-paying roles increase AI usage by 8%, while mid- and low-paying roles decrease by 4% to 11%. Gender analysis shows different collaboration preferences: women are more likely to iterate collaboratively, using command-line tools and automated tasks 6.3 and 7.3 percentage points less, respectively, but they spend more total time chatting with AI, going back and forth.