The January 2026 Anthropic Economic Index is the fourth report in Anthropic's series of first-party usage-data studies. It introduces five "economic primitives" for classifying AI usage (task complexity, skills required, use case, autonomy, and success rate) and reports usage composition, geographic patterns, productivity extrapolations, and labor-market coverage drawn from 1 million Claude.ai conversations and 1 million first-party API records sampled over November 13–20, 2025. The report is provider-side observational data: Anthropic measuring its own traffic using its own classifier (Claude), which shapes both what it reveals and the framing cautions that attach to it.
Metadata
| Field | Value | |
|---|---|---|
| Publisher | [[anthropic | Anthropic]] |
| Lead authors | Ruth Appel, Maxim Massenkoff, Peter McCrory | |
| Contributors | Miles McCain, Ryan Heller, Tyler Neylon, Alex Tamkin | |
| Date | 2026-01-15 | |
| Type | First-party usage-data research report | |
| Data window | November 13–20, 2025 | |
| Sample | 1M Claude.ai conversations + 1M first-party API records | |
| Privacy tooling | CLIO privacy-preserving analysis | |
| Data release | HuggingFace: Anthropic/EconomicIndex |
The report is provider-side observational data, distinct from the third-party productivity and adoption evidence elsewhere in the corpus, which observes a specific sector or surveys users: Brynjolfsson 2023 (customer support), AMA 2026 (physician sentiment), Lancet 2025 (clinical deskilling), and MIT NANDA 2025 (enterprise portfolio). Here the lab measures its own traffic using its own classifier.
The five "economic primitives"
The report's central methodological contribution is to classify every conversation along five dimensions rather than report top-line adoption statistics.
| Primitive | What it measures | Software dev avg | Personal-task avg |
|---|---|---|---|
| Task complexity | Estimated human hours | 3.3 hrs | 1.8 hrs |
| Skills required | Years of education | 13.8 yrs | 9.1–9.4 yrs |
| Use case | Work / Personal / Coursework | — | — |
| Autonomy | 1–5 scale (higher = more autonomous) | ~3.5 | ~3.5 |
| Success rate | Task completed successfully | 61% | 78% |
Personal tasks succeeded more often than software-development tasks (78% versus 61%), an inversion of what a pure-capability view would predict. The report identifies complexity, not domain, as the first-order predictor of success.
Usage composition
Top-line usage composition over the November 13–20, 2025 window differed substantially between the consumer product and the API.
| Dimension | Claude.ai | First-party API |
|---|---|---|
| Top-10 task share | 24% | 32% |
| Computer/Math tasks | 34% | 46% |
| Education tasks | 16% | 4% |
| Work-related use | 46% | 74% |
| Directive mode | 32% | 64% |
| Median task duration (human w/ AI) | 15 min | 5 min |
| Overall success rate | 67% | 49% |
The report characterizes Claude.ai and the API as behaviorally distinct populations rather than the same usage seen at different scales. API usage is more concentrated, more directive, more work-focused, shorter-session, and lower-success per task, a pattern the report reads as production automation versus human-in-the-loop exploration.
Collaboration mode
Augmented use rose to 52% (up 5 percentage points); automation fell to 45% (down 4 percentage points); directive conversations fell to 32% (from 39% in August). Anthropic frames this as users moving back toward collaborative workflows as agent products mature.
Geographic patterns
Within the United States, the Gini coefficient of Claude usage across states fell from 0.37 to 0.32 between August and November 2025; projecting at that rate, the report estimates equalized adoption in 2–5 years. A 1% increase in tech workers is associated with roughly 0.36% higher per-capita Claude usage.
Globally, the report finds no regional convergence. The leading countries by adoption are the United States, India, Japan, the United Kingdom, and South Korea, and adoption correlates with GDP per capita (roughly +0.7% Claude use per +1% GDP per capita). Use case varies by income: higher-income countries skew toward work and personal use, lower-income countries toward coursework. The Balkans and Brazil show the highest work share; Indonesia the highest coursework share.
Productivity estimates
The report includes Anthropic's own back-of-envelope macro translations of observed speedups into productivity-growth contributions.
| Assumption | Unadjusted | Success-adjusted |
|---|---|---|
| Speedup (base, Claude.ai) | 9×–12× | — |
| Annual productivity growth contribution (10-yr) | 1.8 pp | 1.2 pp |
| API data | 1.8 pp | 1.0 pp |
| Complements (σ=0.5) | 0.7–0.9 pp | 0.6–0.8 pp |
| Substitutes (σ=1.5) | 2.2–2.6 pp | — |
These are not empirically measured productivity gains but an extrapolation from observed speedups and task-time estimates, presented as Anthropic's upper-bound aggregation rather than the equivalent of Brynjolfsson's RCT-derived 15%. The extrapolations (1.2–2.6 pp annually) are consistent with moderate Aschenbrenner-style forecasts rather than transformative ones (Source: blog.aifutures.org) / Situational Awareness: The Decade Ahead. They are directionally consistent with Stanford HAI AI Index Report 2026's 14–26% task-level gains but depend on speedup and substitutability assumptions that have not been externally validated.
Labor-market coverage and deskilling
The report finds that 49% of occupations have at least 25% of their tasks covered by Claude usage. The mean predicted education of Claude-covered tasks is 14.4 years, versus 13.2 years across all tasks. It characterizes the net effect as deskilling across most occupations, in the sense that the AI performs the higher-education task within the job. Named examples of deskilling include technical writers, travel agents, and teachers; upskilling, which the report describes as rare, includes real-estate and property managers.
Anthropic's "deskilling" here is a task-composition claim, estimated from the mean years of education associated with Claude-handled tasks versus the full occupation. It is distinct from the Lancet/AMA definition tracked at AI Deskilling, which measures erosion of unaided human performance over time. The Anthropic Economic Index does not measure unaided performance, so it cannot observe Lancet-style deskilling by design. The two senses can coexist, and the Anthropic task-composition pattern is a plausible mechanism for the Lancet-style outcome, but they are different measurements and are not interchangeable. See AI Deskilling for the full disambiguation.
Key claims and confidence
| Claim | Confidence | Notes |
|---|---|---|
| Top-10 Claude.ai task share was 24% in Nov 2025 | high | Direct sample measurement |
| Augmentation (52%) now exceeds automation (45%) on Claude.ai | high | Direct measurement; see AI and Productivity augmentation vs displacement |
| Claude.ai and API are behaviorally distinct populations | high | Strong quantitative separation across 6+ metrics |
| US state adoption converging at Gini −0.05 per 3 months | medium | Short time horizon; March 2026 report revises to slower convergence |
| 1.2–1.8 pp annual productivity contribution | low | Model-based extrapolation from speedups, not measured productivity |
| "Net deskilling" across most occupations | medium | Based on task-embedding education estimates; mechanism differs from AI Deskilling as defined from Lancet Endoscopist Deskilling Study (2025) |
Reading against other sources
Because Anthropic publishes these reports, several framing cautions attach to the data:
- Measurement asymmetry. Success rates are measured by Claude classifying whether Claude succeeded. The report acknowledges classifier-version variance (Sonnet 4.5 versus Sonnet 4). This compares against NANDA's 95% enterprise-pilot-failure finding, which measures P&L impact, a harder bar.
- Denominator control. Anthropic chooses what "task" means, what "success" means, and what share of usage to report. Top-10 concentration metrics compress thousands of tasks into summary statistics whose construction affects the headline.
- Productivity aggregation. The 1.2–1.8 pp annual figure is Anthropic's own extrapolation, not an independent measurement, and depends on speedup and substitutability assumptions that have not been externally validated.
- Composition data. Composition data (coding share, geographic distribution, the Claude.ai versus API behavioral split) is hard to fabricate and is new information.
The report's composition data is provider-measured but verifiable in structure; its success-rate data is provider-graded; and its productivity extrapolations are an Anthropic scenario rather than a measurement.
Relevance to existing content
- AI and Productivity — adds first-party provider-side data complementing Brynjolfsson (task RCT), AMA (sentiment), Lancet (endpoint), and MIT NANDA (enterprise portfolio), under the section "First-Party Deployment-Reality Data."
- Enterprise AI Deployment Gap — provides the provider-side view of what gets through; NANDA provides the enterprise-side view of what doesn't. The two are complementary, not contradictory.
- Agentic AI — the autonomy primitive gives a first quantitative handle on how agentic actual usage is (about 3.5/5 average for both software dev and personal tasks).
- AI Labor Disruption — the deskilling claim and the 49% task-coverage figure are directly relevant.
- AI Deskilling — the Anthropic "deskilling" definition (AI doing the higher-education task within a job) is distinct from the Lancet/AMA definition (erosion of unaided human performance).
Relationships
- supports: AI and Productivity — first-party provider-side evidence complementing task/sentiment/endpoint/enterprise evidence already on the page
- supports: Enterprise AI Deployment Gap — provider-side usage composition paired with NANDA's enterprise-side deployment failure
- supports: Agentic AI — quantitative handle on autonomy and coding share
- related: AI Labor Disruption, AI Deskilling, AI Benchmarks and Evaluation
- related: Anthropic Economic Index — March 2026: Learning Curves — continuation report (March 2026) adding learning-curve findings and showing task-share dropping from 24% → 19%
- related: Stanford HAI AI Index Report 2026 — third-party macro adoption view; similar directional findings, different measurement