Anthropic Economic Index January 2026 Report – Economic Primitives

TL;DR

Anthropic’s January 2026 Economic Index report adds five foundational metrics—task complexity, human/AI skill levels, use case, AI autonomy, and task success—to its analysis of Claude usage, showing that AI adoption remains geographically uneven, that higher‑income regions use Claude more collaboratively, and that current AI success rates temper projected productivity gains.

---

Introduction: New Metrics for AI‑Economic Impact

The report presents economic primitives, simple, privacy‑preserving measures derived by asking Claude to classify anonymized Claude.ai and first‑party (1P) API transcripts. These primitives cover five dimensions relevant to AI’s macroeconomic impact:

1. Task complexity – estimated human‑only time, AI‑assisted time, and multitasking.
2. Human and AI skills – years of education needed to understand prompts and responses.
3. Use case – work, coursework, or personal.
4. AI autonomy – degree of decision‑making delegated to Claude.
5. Task success – Claude’s self‑assessment of whether it completed the task.

The data set includes one million randomly sampled Claude.ai conversations (Free, Pro, Max) and one million 1P API records from 13–20 Nov 2025, the period just before the Opus 4.5 release.

---

1. Changes Since the September 2025 Report

1.1 Task Concentration Persists

• The top 10 tasks account for 24 % of Claude.ai conversations (up from 23 %); for 1P API traffic the share rises to 32 % (up from 28 %).
• Coding‑related tasks dominate: ~34 % of Claude.ai and 46 % of API conversations involve computer‑and‑mathematical work.

1.2 Augmentation Re‑gains Dominance on Claude.ai

• Augmented interactions (task iteration, learning, validation) rise to 52 % of Claude.ai conversations, while automated (directive) use falls to 45 %.
• Product releases—file creation, persistent memory, and workflow “Skills”—appear to have shifted users toward more collaborative patterns.

1.3 Regional Diffusion

• U.S. convergence: Usage per capita is becoming more even across states; the Gini coefficient drops from 0.37 to 0.32. Simple extrapolation suggests parity could be reached in 2–5 years.
• Global concentration: The Anthropic AI Usage Index (AUI) remains tightly linked to GDP per capita; no clear sign of convergence across countries.

---

2. Economic Primitives – Definition, Validation, and Limitations

2.1 Operationalization

Primitive	How Measured	Example Prompt
Task complexity	Claude estimates human‑only time, AI‑assisted time, and whether multiple tasks appear in a single conversation.	“Debug this Python function.”
Human & AI skills	Claude predicts years of formal education required to understand the user’s prompt and its own response.	“Explain the statistical model behind this regression.”
Use case	Classifier tags the conversation as work, coursework, or personal.	“Help me write a literature review for my class.”
AI autonomy	Scale 1–5 based on how much decision‑making Claude performs without user back‑and‑forth.	“Generate a complete marketing plan.”
Task success	Claude rates whether it believes the answer satisfies the request.	“Summarize the key findings of this paper.”

2.2 Validation

• Directional accuracy was confirmed against a small human‑annotated subset of Claude.ai transcripts and synthetic API data.
• For three primitives (human‑only time, AI‑assisted time, AI autonomy) chain‑of‑thought prompting improved classifier performance; for others a simpler prompt sufficed.
• External benchmarks align: the human‑education estimate correlates (r > 0.92) with BLS occupational education levels; time‑estimate predictions match observed productivity gains in Anthropic’s prior work.

2.3 Caveats

• Primitives are noisy and should be interpreted as signals, not exact measurements.
• Sampling thresholds (≥15 conversations, ≥5 users) exclude low‑volume regions and rare tasks.
• Success rates reflect both model capability and user selection bias—users may avoid tasks they expect to fail.

---

3. Geographic Variation in AI Use

3.1 Adoption Patterns by Income

• Work vs. coursework vs. personal: Higher‑GDP per‑capita countries show larger shares of work (e.g., Denmark, US) and personal use, while lower‑income countries exhibit more coursework usage (e.g., Indonesia).
• A 1 % increase in GDP per capita predicts a 0.7 % rise in AUI at the country level.

3.2 U.S. State‑Level Drivers

• States with larger shares of computer & mathematical occupations have higher AUI; a 1 % increase in such workers yields a 0.36 % increase in usage per capita.
• After controlling for GDP, the education‑year primitive loses statistical significance, suggesting workforce composition is the primary driver within the U.S.

3.3 Autonomy and Collaboration

• High‑income regions exhibit lower automation and lower AI autonomy, indicating a collaborative (augmentation) style.
• This pattern is not statistically significant across U.S. states, likely due to narrower income variation.

---

4. Task‑Level Productivity and Labor‑Market Implications

4.1 Speed‑up vs. Success Trade‑off

• Speed‑up (human‑only time ÷ human‑with‑AI time) rises with task education level: ~9× for 12‑year tasks, 12× for 16‑year tasks on Claude.ai.
• Success rate declines for more complex tasks: ~70 % for sub‑high‑school tasks, 66 % for college‑level tasks.
• Adjusting speed‑up by success probability still leaves a positive education gradient.

4.2 Effective AI Coverage of Occupations

• Effective AI coverage = Σ (task‑share × success‑rate). It differs from raw task coverage because some high‑frequency tasks have low success.
• Example: Data entry clerks achieve high effective coverage despite low raw coverage because their dominant task (data entry) has high success.
• Radiologists and medical transcriptionists also show high effective coverage, whereas microbiologists fall below the 45° line due to uncovered lab‑hands‑on work.

4.3 Deskilling vs. Upskilling

• Claude’s observed tasks have a mean predicted education requirement of 14.4 years, about 1 year higher than the economy‑wide average (13.2 years).
• Removing Claude‑covered tasks generally lowers the average education requirement of the remaining work, implying deskilling for occupations such as technical writers, travel agents, and many teaching roles.
• Some occupations (e.g., real‑estate managers) experience upskilling because routine administrative tasks are automated while high‑skill negotiation tasks remain.

4.4 Revised Productivity Estimates

• Baseline (speed‑up only) implied 1.8 pp annual labor‑productivity growth over the next decade.
• Incorporating task success reduces the estimate to 1.2 pp (Claude.ai) and 1.0 pp (API).
• Accounting for task complementarity via a CES aggregation shows the effect is highly sensitive to the elasticity of substitution (σ):
  • σ = 0.5 → 0.7–0.9 pp per year.
  • σ = 1.5 → 2.2–2.6 pp per year.
• The API sample is more sensitive to σ because of its higher task concentration.

---

5. Key Takeaways for Policymakers and Researchers

1. Economic primitives provide a richer, multidimensional view of AI usage than the binary automation/augmentation split.
2. Geographic inequality persists: high‑income countries adopt Claude more intensively and collaboratively, while low‑income regions focus on coursework.
3. Skill bias is evident: Claude is used disproportionately on higher‑education tasks, which may lead to net deskilling of many occupations.
4. Productivity gains are real but moderated by task success rates and the degree of complementarity among tasks.
5. Rapid U.S. diffusion (potential parity in 2–5 years) contrasts with slower global convergence, suggesting domestic policy can influence adoption speed more directly.
6. Human capital matters: the strong correlation (r ≈ 0.93) between user prompt education and Claude’s response complexity underscores the need for education and training to unlock AI benefits.

---

6. Data Availability

Anthropic releases the full anonymized dataset, including primitive classifications, at:

• Hugging Face repository: https://huggingface.co/datasets/Anthropic/EconomicIndex
• Supplementary figures (Figures 1.1–1.8, 2.1–2.2, 3.1–3.5, 4.1–4.6) are hosted on Anthropic’s site and referenced throughout the report.

---

7. Future Directions

• Track how task success evolves with newer models (e.g., Opus 4.5) and whether the effective task horizon expands.
• Extend primitives to capture ethical and trust‑safety dimensions of AI use.
• Combine API and Claude.ai streams to model the transition from interactive chat to production‑grade automation.
• Investigate policy levers that could accelerate equitable diffusion, such as subsidized AI access and targeted AI literacy programs in low‑GDP regions.

---

Report authors: Ruth Appel, Maxim Massenkoff, Peter McCrory, Miles McCain, Ryan Heller, Tyler Neylon, Alex Tamkin.