Anthropic launches economic index with 5 AI 'primitives'

Jan 15, 2026Keryc Díaz5 minutes

Anthropic has published its fourth Anthropic Economic Index with a notable change: five 'economic primitives' that describe not only how much Claude is used but how it's used. Why should you care? Because the way you interact with AI determines which tasks can be automated, which jobs get transformed, and how productivity shifts.

What are the 'economic primitives' and why they matter

The 'primitives' are simple but informative measures obtained by asking Claude to classify anonymous conversations. They capture five dimensions: task complexity, human and AI skills, use case (work, study, or personal), degree of autonomy delegated to the AI, and the task's success rate.

Why is this useful? Knowing how often an AI is used isn't enough. You need to know if the AI handles short or long tasks, whether the prompt needs a lot of education to understand, whether you let the AI make decisions, or whether it fails often. Those differences change the economic impact dramatically.

Key technical findings (summarized)

Concentration of tasks: the 10 most frequent tasks explain 24% of use on Claude.ai and 32% of traffic on 1P API for enterprise customers. A lot of activity remains concentrated in coding and other high-value tasks.
Augmentation vs automation: in November 2025 interaction on Claude.ai returned to being mostly augmentation (52%) versus automation (45%). 1P API is dominated by automation due to its programmatic nature.
Geographic variation: global use is highly correlated with GDP per capita. Within the U.S., states with more workers in computing occupations show higher use; however, lower-use states are growing faster and, by estimates, parity could be reached in 2–5 years if trends hold.
New metrics and validation: the nine final classifiers (implemented with prompts to Claude and validated against humans and synthetic data) are "directionally accurate": not exact, but they capture useful relationships between education, complexity, and success.
Human-AI education correlation: the measure of years of education required to understand the prompt and the response is strongly correlated (r > 0.92). In short: how you write determines how the AI responds.

Times, success and "task horizons"

Complexity and speedup: more complex tasks show larger speedups. For example, prompts that require ~16 years of education achieve greater speedups than prompts requiring ~12 years.
Success vs complexity tradeoff: longer or more complex tasks have lower success rates. In 1P API the fitted intersection indicates ~50% success around 3.5 hours of estimated human work; on Claude.ai the point is much farther (~19 hours), possibly because of the advantage of multi-turn dialogue.
Interpretation: controlled benchmarks (like METR) measure the autonomous limit; real usage data reflects both model capability and users' selection of what tasks they hand to the AI.

Impact on occupations and productivity (technical)

Effective coverage vs simple coverage: they introduce “effective AI coverage,” which weights not only which tasks appear in the data but how much time they occupy in the workday and their success rate. Some occupations (e.g., data entry) show high effective coverage because their tasks are repetitive and have high success rates.
Deskilling vs upskilling: when you remove tasks Claude already covers, the average profile of remaining tasks tends to lower the education required for most occupations (net deskilling). But there are opposite cases: for example, property managers may be left with more skilled tasks if routine work is automated.
Aggregate product and reliability: prior estimates projected ~+1.8 percentage points annual labor productivity growth from current AI adoption. If you incorporate per-task success rates, the figure falls to ~+1.2 pp for Claude.ai and ~+1.0 pp for 1P API.
Task complementarity (CES): using a CES aggregator, the aggregate effect depends critically on the elasticity of substitution σ. If tasks are complementary (σ < 1) bottlenecks sharply reduce gains; if substitutable (σ > 1) gains amplify. Adjusting for success and complementarity, estimates can fall to ~0.6–0.9 pp or rise above 2 pp depending on σ.

Concrete examples (to make it clear)

Data entry: even if few tasks appear in the data, they occupy a lot of time and Claude has a high success rate, so effective coverage is high and the job can be substantially automated.
Travel agents: they could experience deskilling, because the AI covers complex planning and leaves more routine tasks to humans.
Property managers: they might be left with more strategic work and therefore see an increase in the average skill level of their tasks.

Limitations and how to read this report

Sample and selection: figures come from 1M conversations on Claude.ai and 1M 1P API transcriptions from a specific period (13–20 Nov 2025). Users and customers select tasks. This is not a controlled experiment.
Classifiers aren't perfect: they're useful for directional signals but shouldn't be read as exact measurements of human time or absolute success.
Uncertainty in diffusion: U.S. internal convergence estimates (β̂ ≈ 0.77 OLS; 2SLS ≈ 0.86–0.89) imply rapid diffusion, but they're based on only three months of change and may attenuate.

What do you take away from this?

How you use AI matters as much as how much you use it. Measuring complexity, autonomy, and success rate helps estimate what can actually be automated.
Today, Claude tends to work on higher-education tasks and offers large speedups, but reliability is lower on the longest, most complex tasks.
Labor implications are heterogeneous: some roles are partially automated, some deskill, and others specialize. Public policy and training will be crucial to ensure AI adoption promotes inclusion and productivity.
Productivity numbers remain large but are sensitive to reliability and how tasks combine within jobs.

Anthropic's contribution is useful: it provides reproducible data and new indicators so academics, governments, and companies can stop guessing and start measuring the labor transitions that are already happening.

Original source

https://www.anthropic.com/research/anthropic-economic-index-january-2026-report

Stay up to date!

Get AI news, tool launches, and innovative products straight to your inbox. Everything clear and useful.

What are the 'economic primitives' and why they matter

Key technical findings (summarized)

Concentration of tasks: the 10 most frequent tasks explain 24% of use on Claude.ai and 32% of traffic on 1P API for enterprise customers. A lot of activity remains concentrated in coding and other high-value tasks.

Augmentation vs automation: in November 2025 interaction on Claude.ai returned to being mostly augmentation (52%) versus automation (45%). 1P API is dominated by automation due to its programmatic nature.

Geographic variation: global use is highly correlated with GDP per capita. Within the U.S., states with more workers in computing occupations show higher use; however, lower-use states are growing faster and, by estimates, parity could be reached in 2–5 years if trends hold.

New metrics and validation: the nine final classifiers (implemented with prompts to Claude and validated against humans and synthetic data) are "directionally accurate": not exact, but they capture useful relationships between education, complexity, and success.

Human-AI education correlation: the measure of years of education required to understand the prompt and the response is strongly correlated (r > 0.92). In short: how you write determines how the AI responds.

Times, success and "task horizons"

Complexity and speedup: more complex tasks show larger speedups. For example, prompts that require ~16 years of education achieve greater speedups than prompts requiring ~12 years.

Success vs complexity tradeoff: longer or more complex tasks have lower success rates. In 1P API the fitted intersection indicates ~50% success around 3.5 hours of estimated human work; on Claude.ai the point is much farther (~19 hours), possibly because of the advantage of multi-turn dialogue.

Interpretation: controlled benchmarks (like METR) measure the autonomous limit; real usage data reflects both model capability and users' selection of what tasks they hand to the AI.

Impact on occupations and productivity (technical)

Effective coverage vs simple coverage: they introduce “effective AI coverage,” which weights not only which tasks appear in the data but how much time they occupy in the workday and their success rate. Some occupations (e.g., data entry) show high effective coverage because their tasks are repetitive and have high success rates.

Deskilling vs upskilling: when you remove tasks Claude already covers, the average profile of remaining tasks tends to lower the education required for most occupations (net deskilling). But there are opposite cases: for example, property managers may be left with more skilled tasks if routine work is automated.

Aggregate product and reliability: prior estimates projected ~+1.8 percentage points annual labor productivity growth from current AI adoption. If you incorporate per-task success rates, the figure falls to ~+1.2 pp for Claude.ai and ~+1.0 pp for 1P API.

Task complementarity (CES): using a CES aggregator, the aggregate effect depends critically on the elasticity of substitution σ. If tasks are complementary (σ < 1) bottlenecks sharply reduce gains; if substitutable (σ > 1) gains amplify. Adjusting for success and complementarity, estimates can fall to ~0.6–0.9 pp or rise above 2 pp depending on σ.

Concrete examples (to make it clear)

Data entry: even if few tasks appear in the data, they occupy a lot of time and Claude has a high success rate, so effective coverage is high and the job can be substantially automated.

Travel agents: they could experience deskilling, because the AI covers complex planning and leaves more routine tasks to humans.

Property managers: they might be left with more strategic work and therefore see an increase in the average skill level of their tasks.

Limitations and how to read this report

Sample and selection: figures come from 1M conversations on Claude.ai and 1M 1P API transcriptions from a specific period (13–20 Nov 2025). Users and customers select tasks. This is not a controlled experiment.

Classifiers aren't perfect: they're useful for directional signals but shouldn't be read as exact measurements of human time or absolute success.

Uncertainty in diffusion: U.S. internal convergence estimates (β̂ ≈ 0.77 OLS; 2SLS ≈ 0.86–0.89) imply rapid diffusion, but they're based on only three months of change and may attenuate.

What do you take away from this?

How you use AI matters as much as how much you use it. Measuring complexity, autonomy, and success rate helps estimate what can actually be automated.

Today, Claude tends to work on higher-education tasks and offers large speedups, but reliability is lower on the longest, most complex tasks.

Labor implications are heterogeneous: some roles are partially automated, some deskill, and others specialize. Public policy and training will be crucial to ensure AI adoption promotes inclusion and productivity.

Productivity numbers remain large but are sensitive to reliability and how tasks combine within jobs.