The result: five new "economic primitives." These are simple metrics that show how people actually use AI. Not how tech companies claim it is used.

The data covers five dimensions: user and AI capabilities, task complexity, degree of autonomy, success rate, and use case (work, study, personal).

Key Findings

A lot has happened since the last report in September 2025. But not everything.

The Five New Primitives

Anthropic introduces five new metrics. They capture what is actually happening in AI usage:

These primitives are not perfect. But they point the way forward. Anthropic validated them against external benchmarks. For example: estimated years of education strongly correlate with actual worker education levels across different occupations.

The goal: provide signals. Not definitive truths. Multiple simple metrics taken together provide a clearer picture than a single complex, potentially flawed measurement.

What the Data Shows: Geographic Differences

The primitives reveal striking differences. Between countries. And within the US.

GDP drives usage. A one percent increase in GDP per capita translates to 0.7% more Claude usage. This relationship holds globally and across the US. Wealthy countries use AI differently: more for personal projects. Lower-income countries: emphasis on study and specific applications (e.g., coding).

Education correlates with usage. Countries and US states where users write more complex prompts (requiring higher educational levels) use Claude more intensively. However: within the US, this effect disappears when controlling for GDP. There, education serves more as a proxy for economic development.

Success rates vary paradoxically. Globally: higher user education correlates with a lower success rate. Why? More educated users ask harder questions. Within the US: the opposite pattern (higher education, higher success rate). However, this effect disappears when controlling for other factors.

Input dictates output. A central finding: the years of education required to understand the user prompt correlate almost perfectly with the years of education needed for Claude's response (r > 0.92). In other words: sophisticated questions yield sophisticated answers. Simple questions yield simple answers.

Tasks and Productivity: The Trade-offs

The more complex the task, the greater the time savings. But also: the lower the success rate.

Tasks requiring 12 years of education (high school level) achieve a 9-fold speedup. Tasks requiring 16 years (bachelor's level)? A 12-fold speedup. This means: AI provides greater assistance with complex, highly skilled activities.

However, Claude fails more frequently on difficult tasks. On simple tasks, the success rate is 70%. On complex ones, it drops to 66%. Even so, the net gain remains higher for complex tasks.

Task horizons in practice. Task duration affects the success rate. With API usage, it drops from 60% (under 1 hour) to 45% (over 5 hours). The 50% threshold lies at 3.5 hours. On Claude.ai, it takes longer for the success rate to fall: 19 hours. Why? Multi-turn conversations allow for corrections. Users break complex tasks down into steps.

Job exposure recalculated. Previously, Anthropic only measured the percentage of a job covered by AI. Now: "Effective AI Coverage." This weights exposure by success rate and the proportion of time dedicated to each task.

Data entry clerks show high coverage. Why? Their main task (entering data from documents) has a high success rate with Claude. Microbiologists? Low coverage. While AI covers many of their tasks, it does not cover the time-intensive ones (laboratory work).

Deskilling or Upskilling?

When AI takes over certain tasks, job content shifts. But in which direction?

Anthropic analyzed: which tasks does Claude take over on average? The answer: those requiring higher educational levels. The average of all tasks across the economy: 13.2 years of education. Tasks performed by Claude: 14.4 years.

This means: when Claude strips away these tasks, less demanding tasks remain on average. The net effect: deskilling.

Example of travel agency staff: Claude takes over itinerary planning and cost estimation (higher education). What remains: printing tickets, processing payments (lower education). Deskilling.

Counterexample of real estate managers: Claude takes over routine administrative tasks (record-keeping, market comparisons). What remains: loan negotiations, stakeholder management (higher education). Upskilling.

These patterns are not set in stone. They are based on current usage. As new AI capabilities emerge, the affected tasks will change.

Productivity Gains: The Reality

Earlier estimates: AI could increase labor productivity by 1.8 percentage points per year. Over a decade.

Now, weighted by success rates: the effect is cut in half. To approximately 1.0 percentage point per year. This accounts for the fact that Claude does not complete all tasks perfectly. Validation takes time.

Yet even 1.0 percentage point is economically significant. That would represent a return to the productivity growth rates of the late 1990s and early 2000s.

Task complementarity as a bottleneck. If certain tasks are essential and cannot be substituted, it constrains productivity. Teachers can create lesson plans faster with AI. But classroom time remains the same.

Anthropic models various scenarios (CES aggregation). With strong complementarity, the productivity gain drops to 0.6–0.8 percentage points. With strong substitutability, it rises to 2.2–2.6 percentage points.