This post is the short public summary of the fourth Anthropic Economic Index report, released the same day. The full report is covered in Anthropic Economic Index report: Economic primitives, which has the methods, caveats and chapter-level locations; this note records what the summary chose to emphasize and how it framed the results. The post opens with three questions: whether AI is making people faster at work, which tasks it supports best, and how it might change occupations. Its answer is a new layer of measurement, five "economic primitives" that Claude assigns to every sampled conversation, which the post describes as "a leading indicator of AI's potential economic impacts". The page names its team (Economics) but no individual authors.
What the post describes
The sample is conversations from November 2025, "predominantly using Claude Sonnet 4.5", taken from Claude.ai (mostly consumer use) and from Anthropic's first-party API (mostly business use). Footnote 1 gives the sizes: one million conversations from each source, analyzed with the privacy-preserving method of Clio: Privacy-Preserving Insights into Real-World AI Use. The five primitives, defined in footnote 2, are task complexity (estimated human time with and without AI, and multitasking), human and AI skills (whether the user could have done the task alone, and the years of education needed to understand the prompt and the response), use case (work, education or personal), AI autonomy (how much decision-making the user delegates), and task success (Claude's assessment of whether it completed the task). The post points readers to chapter two of the report for how the primitives were tested.
What it reports
The post groups results into tasks, occupations and aggregate impact. All numbers below are as the post states them; the report gives the same figures with more detail.
Speedups grow with task complexity. On Claude.ai, tasks whose prompts need a high-school education (12 years) were sped up about 9 times, and those needing a college degree (16 years) about 12 times, with larger speedups on the API. Success falls slightly with complexity, from 70% for tasks needing less than high school to 66% for college-level tasks, which "reduces, but doesn't eliminate" the advantage of complex tasks.
Real-world task horizons. METR's benchmark puts Claude Sonnet 4.5 at 50% success on tasks of about 2 hours. In Anthropic's data, the 50% point falls at about 3.5 hours for the API and about 19 hours for Claude.ai. The post suggests the gap "might not be as discordant as it seems" because the methods differ: users can split complex tasks into steps and correct course, and they bring Claude tasks "they're more confident will work", a form of selection bias.
Use differs with national income. Higher-GDP countries use Claude more for work and personal purposes, lower-income countries more for coursework, which the post calls an "adoption curve" story and connects to a Microsoft study and to an Anthropic program in Rwanda with the training provider ALX.
Coverage weighted by success. The share of jobs with Claude used for at least a quarter of their tasks was 36% in the first report (January 2025 data) and is 49% when data are pooled across reports. Weighting tasks by time, frequency and success makes some occupations, such as data entry keyers and radiologists, look more affected than coverage alone suggests, and others, such as teachers and software developers, less.
Covered tasks need more education. Tasks Claude covers require an average of 14.4 years of education against 13.2 for the economy as a whole. Removing them would, as a first-order effect, deskill jobs on average (technical writers, travel agents and teachers among them), while a few occupations such as real estate managers would move the other way. The post says that this is "not necessarily predicting" deskilling, since labor markets adjust and covered tasks will change as models improve.
Productivity estimate reduced. Speedups alone reproduce Anthropic's earlier estimate of 1.8 percentage points a year of additional US labor productivity growth over ten years. Weighting time savings by the probability of success lowers it by about a third for Claude.ai, to 1.2 points, and to 1.0 point for the API. The post adds that even 1 point would return US productivity growth to late-1990s and early-2000s rates, and that the estimate does not account for more capable models such as Claude Opus 4.5.
Updates to earlier measures. The top ten tasks hold 24% of Claude.ai conversations, up from 21% in January 2025. Computer and mathematical tasks are about a third of Claude.ai conversations and nearly half of API traffic. Augmentation (52%) has overtaken automation (45%) on Claude.ai, reversing the August 2025 sample (49% automation against 47% augmentation); the post gives the January 2025 split as 55% to 41% and March 2025 as 55% to 42%, so over the year automation's share has still risen slowly. (The full report gives the January 2025 augmentation share as 56%.) Geographic concentration persists, but usage across US states is evening out, and the report's model predicts equalization within two to five years if the trend holds.
How the post frames the results
The conclusion's main message is unevenness: AI use remains concentrated by country and occupation, and affects occupations differently. The post presents the primitives as a baseline for future reports and names two things it expects to watch: Claude taking on harder tasks with more success, and tasks moving from Claude.ai to the API as they become reliable, which it would read as a sign of business adoption. Compared with the full report, the summary leaves out the classifier validation details, the geographic regressions, the diffusion model's estimates and uncertainty, and the task-substitutability scenarios that move the productivity estimate between 0.6 and 2.6 points.
Limitations
The post states two limits directly: its success-adjusted coverage measure covers only tasks seen on Claude.ai, and "it's not always clear" how a conversation maps to real-world change; and the time-horizon comparison with METR mixes different methods. The full report's limitations (model-judged success, directional rather than exact classifiers, selection effects, single-turn API records) apply here as well; see Anthropic Economic Index report: Economic primitives.
Research questions
- How much of the 3.5-hour and 19-hour horizons is selection? The post names users' choice of tasks as a source of bias. Comparing success-versus-duration curves for users who bring many varied tasks with those who bring few would indicate how far selection inflates the horizon.
- Does the one-third reduction in the productivity estimate hold under other reliability adjustments? The post multiplies time savings by success probability. Alternatives, such as charging failed attempts their full time, would show how sensitive the 1.2 and 1.0 figures are to that choice.
- Do the occupations the post singles out (data entry keyers, radiologists, teachers, software developers) show labor-market changes in the predicted direction? Matching the effective-coverage ranking against later employment data would test the measure.
- Will tasks actually migrate from Claude.ai to the API as success rises? The post proposes this as an indicator; tracking task shares by platform alongside success rates across future reports would answer it.
How it connects
- Full report. Anthropic Economic Index report: Economic primitives.
- Earlier work the post cites. The first report (The Anthropic Economic Index), the software-development report (Anthropic Economic Index: AI’s impact on software development), the geographic report (Anthropic Economic Index report: Uneven geographic and enterprise AI adoption) and the productivity estimate (Estimating AI productivity gains from Claude conversations).
- Later posts that cite it. The Anthropic Education Report: The AI Fluency Index links this post when describing augmentative use.
References
- Anthropic. (2026). Anthropic Economic Index: New building blocks for understanding AI use. 15 January 2026. https://www.anthropic.com/research/economic-index-primitives (read on anthropic.com; the vault folder for this item holds the full report rather than this post).
- Appel, R., Massenkoff, M., McCrory, P., et al. (2026). Anthropic Economic Index report: Economic primitives. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report.