Anthropic Economic Index report: Economic primitives

economic indexdeployment datalabor marketsproductivitytask successAI autonomyprivacy-preserving analysiseconomic research

The fourth Anthropic Economic Index report samples one million Claude.ai conversations and one million first-party API transcripts from November 2025, adds five "economic primitives" (task complexity, human and AI skills, use case, AI autonomy and task success) classified by Claude, and uses them to study geographic adoption, real-world task horizons, success-weighted job coverage, the skill content of covered tasks, and a revised productivity estimate that falls from 1.8 to 1.2 (Claude.ai) or 1.0 (API) percentage points a year once task success is factored in.

Ruth Appel, Maxim Massenkoff, Peter McCrory et al. (Anthropic) · 2026-01-15 · original

Earlier Anthropic Economic Index reports counted which occupational tasks people bring to Claude and whether they automate those tasks or work through them with Claude. Those counts say little about how hard the tasks are, how much the user leaves to the model, or whether the model actually succeeds. The fourth report, published on 15 January 2026 by Anthropic's Economic Research team, adds measures for those questions. It calls them "economic primitives", defined in the Introduction as "simple, foundational measures of how Claude is used", and generates them by asking Claude a fixed set of questions about each sampled transcript. The report then uses the primitives to revisit three questions from earlier reports: where Claude is adopted, which jobs it touches, and what its use implies for aggregate labor productivity.

The data come from November 2025, before the release of Claude Opus 4.5 (Introduction). The report was released with a public dataset and an online appendix that prints the classification prompts. Anthropic published a short summary post the same day, covered in Anthropic Economic Index: New building blocks for understanding AI use.

Terms used in this note

  • O*NET task. A task statement from the US Department of Labor's O*NET database, which describes occupations as lists of tasks (for example, "modifying software to correct errors"). Conversations are mapped to these tasks, and tasks to occupations.
  • Automation and augmentation. The two categories into which the report groups its five collaboration patterns (Chapter 1, footnote 5). Automation covers directive use (the user hands over a task, minimal back-and-forth) and feedback loops. Augmentation covers learning, task iteration and validation (asking Claude to check one's own work).
  • 1P API. First-party API traffic, the requests developers and businesses send directly to Anthropic's API. The report calls it "enterprise data"; Claude.ai traffic is "consumer data" (Chapter 1, footnote 1).
  • Anthropic AI Usage Index (AUI). A geography's share of Claude usage divided by its share of the working-age population. An AUI above 1 means more use than population alone predicts; Denmark's AUI of 2.1, for example, means roughly twice the expected rate (Chapter 1).
  • Speedup. Claude's estimate of how long a task would take a competent professional without AI, divided by its estimate of how long the user spent with AI. Reducing a one-hour task to ten minutes is a 6x speedup (Chapter 4).
  • Effective AI coverage. The share of a worker's time that Claude could perform successfully: a weighted sum of task success rates, each task weighted by its share of working time and its frequency (Chapter 4).

What the report does

Data

Both samples come from 13 to 20 November 2025: one million Claude.ai Free, Pro and Max conversations and one million 1P API transcripts (Chapter 1, footnote 1). For the API, each record is a single prompt and response, sometimes taken from the middle of a longer session, because API requests arrive without metadata linking them to earlier exchanges (Chapter 2). The analysis runs through Anthropic's privacy-preserving pipeline, described in Clio: Privacy-Preserving Insights into Real-World AI Use, which drops any cell (a country, or a country and task pair) with fewer than 15 conversations or 5 unique accounts, and applies a stricter floor of 500 conversations and 250 accounts to bottom-up request clusters (Chapter 3, footnote 1). Countries are assigned from IP addresses; VPN and hosting traffic is excluded, as are the Seychelles and Wyoming because much of their traffic in the sampling window was abusive (Chapter 3, footnotes 2, 5 and 6). Classification used Claude Sonnet 4.5, where the previous report had used Sonnet 4 (Chapter 1, footnote 7).

The five primitives

The report adds five primitives to the existing collaboration-pattern measure (Chapter 2, Table 2.1):

  1. Task complexity, measured by estimated human time without AI, estimated time with AI, and whether the user worked on several tasks in one conversation.
  2. Human and AI skills, measured by whether the user could have completed the task alone and by the years of formal education needed to understand the user's prompts and, separately, Claude's responses.
  3. Use case: work, coursework or personal.
  4. AI autonomy: how much decision-making the user leaves to Claude, on a 1 to 5 scale. The report separates this from automation: "Translate this paragraph into French" is directive, hence automation, but leaves Claude little to decide (Chapter 2).
  5. Task success: Claude's judgment of whether it completed the user's task.

Selection and validation

The team chose dimensions for expected economic relevance, complementarity, and whether Claude could classify them with "directional accuracy" (Chapter 2). New classifiers were checked against a human researcher on Claude.ai transcripts that users had submitted with feedback, and on internal and synthetic data for the API. Classifiers that performed poorly were revised or dropped; for task success a simple classifier beat a more nuanced one against human ratings, and chain-of-thought prompting was kept only for the two time estimates and AI autonomy, where it "substantially improved performance" (Chapter 2). Nine new classifiers remained. The report gives no accuracy figures for them. Its stated position is that the measures are directionally correct but not exact, and it names one known bias: Claude "may somewhat underestimate the human education years needed for many tasks" (Chapter 2). As external checks it cites its earlier finding that Claude's time estimates track real software-engineering task times, and shows that its education measure correlates with the share of workers holding a bachelor's degree across occupations (Figure 2.1).

What it finds

Chapter 1: concentration, collaboration and diffusion

  • Concentration across tasks. The ten most common O*NET tasks took 24% of Claude.ai conversations in November 2025, up from 23% in August and 21% in January 2025; on the API they took 32%, up from 28%. The single most common task on both platforms was modifying software to correct errors: 6% of Claude.ai usage and one in ten API records (Chapter 1, Figure 1.1).
  • Coding's share. Computer and mathematical tasks were 34% of Claude.ai conversations, down from a peak of 40% in March 2025, and 46% of API traffic, up from 44% in August. Educational instruction rose on Claude.ai from 9% of conversations in January 2025 to 15%, and office and administrative support rose on the API to 13% (Chapter 1, Figure 1.2).
  • Augmentation back in the lead on Claude.ai. Augmented use rose 5 percentage points to 52% and automated use fell 4 points to 45% (Chapter 1). Directive conversations, 27% in January 2025 and 39% in August, fell to 32%. The report reads the August peak as having "overstated" how fast automation was arriving, while noting that automation was still above its January 2025 level, and it suggests product changes (file creation, memory, Skills) may have pushed use toward collaboration. On the API, about three quarters of interactions were automation (Chapter 2).
  • Global concentration unchanged, US states converging. Across countries the AUI's concentration was essentially unchanged between August and November (Figure 1.6). Across US states the Gini coefficient of usage fell from 0.37 to 0.32. The top five states held nearly half of usage with 38% of the working-age population. Each 1% increase in a state's share of computer and mathematical workers was associated with 0.36% higher usage per capita, which accounts for nearly two thirds of the cross-state variation (Figure 1.7).
  • A diffusion estimate. Modeling each state's AUI as converging proportionally toward 1, the report estimates the convergence parameter at about 0.77 by ordinary least squares and 0.76 weighted, and at 0.89 and 0.86 by two-stage least squares, using workforce composition as an instrument to correct for sampling noise (Chapter 1, Figure 1.8). These imply parity across states within two to five years, which the report compares with roughly 50 years for earlier economically important technologies. It cautions that the estimate rests on three months of data, that only the unweighted 2SLS estimate is distinguishable from 1 at the 10% level, and that "much slower" diffusion cannot be ruled out.

Chapter 2: what the primitives look like

The report contrasts two request clusters with the global average (Figure 2.2). Software development requests are estimated at 3.3 hours of professional time without AI (global average 3.1 hours) and about 13.8 years of education to understand; personal life management requests at 1.8 hours and 9.1 to 9.4 years. Users could have done personal tasks alone 96% of the time and software tasks 82% of the time. Claude judged personal tasks successful 78% of the time and software tasks 61%. Both clusters sat near 3.5 on the 1 to 5 autonomy scale. Across Claude.ai as a whole, use was 46% work, 19% coursework and 35% personal.

Claude.ai and API traffic differ on nearly every primitive (Chapter 2): the API was 74% work against 46%, 64% directive against 32%, and showed task success of 49% against 67%, with shorter tasks (1.7 against 3.1 hours of human time; 5 against 15 minutes with AI). The report notes that part of the success gap may come from Claude.ai's multi-turn format, in which users can clarify and correct.

Chapter 3: geography

  • Use case varies with income. Work and personal use are more common in higher-income countries and coursework in lower-income ones (Figure 3.2), which the report connects to an adoption curve and to a Microsoft report with a similar finding. Indonesia had the highest coursework share; the Balkans and Brazil the highest relative work share (Figure 3.1).
  • Income and adoption. A 1% increase in GDP per capita is associated with a 0.7% increase in Claude usage per capita across countries (Chapter 3). Prompt education correlates positively with the AUI for countries and for US states, but within the US that relationship disappears once GDP and other primitives are controlled for.
  • Relationships that change sign. Task success is negatively associated with prompt education across countries but positively within US states; the state-level relationship becomes insignificant with controls (Figure 3.5). The report's reading is that educated populations across countries may attempt harder tasks.
  • Prompts and responses match. The education level of prompts and of Claude's responses correlate at r = 0.925 across 117 countries and r = 0.928 across 50 US states (Chapter 3).
  • Higher-usage countries collaborate more. Countries with higher usage per capita show less automation and give Claude less autonomy; this is not significant across US states (Chapter 3).

The report stresses that these associations are not causal and that many of the variables are correlated with each other (Chapter 3).

Chapter 4: tasks, horizons, jobs and productivity

Speedup and success trade off. On Claude.ai, tasks whose prompts need 12 years of schooling show about a 9x speedup and those needing 16 years about 12x; speedups are higher on the API throughout (Figure 4.1a). Success falls with complexity, from 70% for tasks needing less than a high-school education to 66% for college-level tasks (Figure 4.1b), but discounting speedups by success does not remove the education gradient. The automation share is essentially unrelated to prompt education (cited in the report as Appendix Figure A.1; Figure 1 of the online appendix).

Real-world task horizons. On the API, success falls from about 60% for tasks under an hour of human time to about 45% for tasks of five or more hours, and the fitted line crosses 50% at 3.5 hours; METR's software benchmark gives 2 hours for Sonnet 4.5 and about 5 hours for Opus 4.5 (Figure 4.3). On Claude.ai the decline is much flatter and the linear fit reaches 50% at about 19 hours. The report attributes part of this to multi-turn conversation breaking tasks into steps, and gives a selection caveat: users choose which tasks to bring, so "observed success rates reflect not just model capability but also user judgment about what will work" (Chapter 4). For the same reason, it says, a better model need not show improvement on this plot. It calls its measure an effective task horizon, a mix of model capability and user behavior.

Job coverage weighted by success. The report's earlier work had found 36% of jobs with Claude usage for at least a quarter of their tasks; pooling reports, the figure is now 49% (Chapter 4). Weighting by time, frequency and success changes which jobs look affected (Figure 4.4). Data entry keyers rank high because, although only two of their nine tasks are covered, the largest one, reading and entering data from source documents, has high success. Radiologists and medical transcriptionists also rise, while microbiologists fall because their most time-consuming work is hands-on lab research. Occupations with near-complete task coverage generally fall below the 45-degree line. The report notes the measure's limit: a conversation mapped to a teacher's lecture task does not mean lecture time on the job falls.

Skill content of covered tasks. Predicting each task's education requirement from its text (a ridge regression on sentence embeddings, Chapter 4, footnote 4), the mean for all tasks in the economy is 13.2 years and for tasks seen in Claude data about 14.4 years (Figure 4.5). Removing covered tasks would therefore deskill most jobs as a first-order effect. Technical writers and travel agents lose their higher-education tasks; real estate managers lose routine record-keeping and are left with negotiation and board work, an upskilling. The report frames this as a simple task-removal exercise inspired by Autor and Thompson (2025), notes that its education measure differs from their concept of expertise, and says the patterns rest on current usage, which will shift as models and applications change.

Aggregate productivity. Using tasks with at least 200 observations, the report replicates its earlier estimate that current usage implies about 1.8 percentage points a year of additional US labor productivity growth over the next decade, for both Claude.ai and API samples (Chapter 4). Multiplying task-level time savings by task success lowers this to 1.2 points for Claude.ai and 1.0 for the API. (The chapter's introduction summarizes the adjustment as roughly halving the gains, "from 1.8 to about 1.0"; the detailed section gives 1.2 for Claude.ai and 1.0 for the API.) A second exercise aggregates tasks within each occupation with a constant-elasticity-of-substitution function (Figure 4.6). At an elasticity of 0.5, where tasks are complements and slow tasks become bottlenecks, the estimate falls to 0.7 to 0.9 points, and to 0.8 (Claude.ai) and 0.6 (API) with the success adjustment; at 1.5 it rises to 2.2 to 2.6 points. Footnote 6 shows how sensitive the baseline is to the task threshold: counting every task above the privacy floor would mechanically raise the estimate to roughly 5 points a year.

The prompts used to classify conversations

The report has no constitution or training principles. Its behavior-shaping text is the classification prompts, printed in the online appendix (Section 2), and every primitive is whatever Claude answers to them. Most are short. The task-success prompt asks only "Did the Assistant complete the task provided by the User successfully?", with Yes or No as the answers (Section 2.1.4). The two education prompts ask how many years of formal education someone would need to understand the user's prompts, or the assistant's responses, on a 0 to 20 scale (Sections 2.1.5 and 2.1.6). The autonomy prompt asks Claude to rate, after two to five sentences of reasoning, how much autonomy the assistant had "to make decisions in this conversation", from 1 (none) to 5 (extreme), with no further definition of the levels (Section 2.2.1). The human-time prompt is more specified: it asks for the hours "a competent professional" would need, assuming domain knowledge, full context, the needed tools and no AI (Section 2.2.2). API versions of the prompts are adapted to single API calls (Section 2.3).

Two points from the report bear on how such wording matters. The validation found that a simpler task-success prompt agreed with human ratings better than a more nuanced one (Chapter 2). And the report reads its near-perfect correlation between the education level of prompts and of responses, partly as a fact about the model's training and instructions: "While Claude is able to respond in a highly sophisticated manner, it tends to do so only when users input sophisticated prompts" (Chapter 3). It contrasts this with a hypothetical model whose system prompt fixes the reading level of every response.

Limitations and open questions

The report states these:

  • Classifier accuracy. The primitives are judged directionally accurate rather than exact, and Claude may underestimate education requirements (Chapter 2). Validation used feedback transcripts and synthetic data, neither fully representative.
  • Model-judged success. Task success is Claude's own assessment of Claude's work.
  • Selection. Users choose which tasks to bring, so success rates and task horizons mix model capability with user judgment (Chapter 4).
  • API format. API records are single prompt-response pairs, so API primitives describe isolated exchanges (Chapter 2).
  • Short window for diffusion. The US convergence estimate rests on three months and may overstate the pace (Chapter 1).
  • Correlation, not causation. Geographic associations do not show what drives adoption (Chapter 3).
  • Coverage is not displacement. It is unclear how often a conversation mapped to a task changes the work done on the job (Chapter 4).

Two further points follow from the report's own details. The headline productivity estimate depends on the task-inclusion threshold (Chapter 4, footnote 6), and the change in classifier model from Sonnet 4 to Sonnet 4.5 between reports, which the report notes "can generate different classification outcomes" in modest ways (Chapter 1, footnote 7), complicates comparisons over time such as the swing between automation and augmentation.

Research questions

  1. How accurate are the nine new classifiers against human raters? The report states directional accuracy but publishes no agreement figures (Chapter 2). Releasing per-classifier agreement on a labeled sample would let users of the dataset weigh each primitive.
  2. How much does the task-success prompt's wording drive the success rates? The success question is a bare Yes/No (online appendix, Section 2.1.4), and a simpler prompt outperformed a nuanced one. Rerunning the classifier with alternative definitions of success on the same transcripts would show how stable figures like the 67% and 49% are.
  3. Is the 19-hour Claude.ai horizon a property of the model or of how users split tasks? The report suggests multi-turn decomposition and selection (Chapter 4). Comparing horizons for single-turn and multi-turn Claude.ai conversations on matched O*NET tasks would separate the two.
  4. Does the education match between prompts and responses come from the user or the model? The r = 0.925 correlation is observational (Chapter 3). An experiment varying prompt sophistication for fixed tasks, and varying a system instruction on response level, would show how much of the match the model's training produces.
  5. Did the August-to-November swing back to augmentation reflect users or the classifier? The classifier model changed between reports (Chapter 1, footnote 7). Reclassifying the August sample with Sonnet 4.5 would measure how much of the 5-point swing remains.
  6. Do success-weighted coverage scores predict labor-market outcomes? The report says the strongest validation will be predictive (Chapter 2). Linking effective coverage by occupation to later employment and wage data would test it.

How it connects

References

  • Appel, R., Massenkoff, M., McCrory, P., McCain, M., Heller, R., Neylon, T., and Tamkin, A. (2026). Anthropic Economic Index report: Economic primitives. Anthropic, 15 January 2026. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report. PDF: https://www-cdn.anthropic.com/096d94c1a91c6480806d8f24b2344c7e2a4bc666.pdf (chapter, figure and footnote numbers in this note refer to the report).
  • Anthropic. (2026). Online Appendix: Economic Primitives and Classification Prompts. In the EconomicIndex dataset, release_2026_01_15/aei_v4_appendix.pdf. https://huggingface.co/datasets/Anthropic/EconomicIndex (read for the prompt wording quoted above).
  • Works cited above only as the report describes them (Autor and Thompson, 2025; Kalanyi et al., 2025; Kwa et al., 2025 (METR); the Microsoft New Future of Work report, 2025) are cited and linked in the report and were not read for this note.