The March 2026 Anthropic Economic Index report ("Learning Curves"), published by Anthropic on 2026-03-24, is a first-party usage-data study of how Claude is deployed. It is the second data point in Anthropic's Economic Index series after the January 2026 report, using a three-month re-sample to document changes in deployment composition over short time horizons and introducing a user-tenure (learning-curve) analysis. The report finds that Claude.ai task usage is diversifying (top-10 task share fell from 24% to 19% between November 2025 and February 2026) while API usage is concentrating, that coding accounts for 35% of Claude.ai conversations, that US per-capita usage is geographically deconcentrating, and that users with longer tenure show higher task success rates after statistical controls.
Provenance
| Field | Value | |
|---|---|---|
| Publisher | [[anthropic | Anthropic]] |
| Lead authors | Maxim Massenkoff, Eva Lyubich, Peter McCrory | |
| Contributors | Ruth Appel, Ryan Heller | |
| Date | 2026-03-24 | |
| Type | First-party usage-data research report | |
| Data window | February 5–12, 2026 | |
| Sample | 1M Claude.ai conversations + 1M first-party API records | |
| Privacy tooling | CLIO privacy-preserving analysis | |
| Occupational taxonomy | O*NET 2019 SOC codes | |
| Wage baseline | BLS OEWS May 2024 |
The January 2026 report established the measurement framework (five economic primitives). The March report applies the same methodology to a later sample, making the index a trend series rather than a single snapshot, and adds the first published user-tenure analysis in the series.
Deployment composition over three months
Comparing the November 2025 and February 2026 samples, the report records the following shifts:
| Metric | Nov 2025 | Feb 2026 | Direction |
|---|---|---|---|
| Top-10 task share (Claude.ai) | 24% | 19% | Diversifying |
| Top-10 task share (API) | 28% | 33% | Concentrating |
| Coursework share (Claude.ai) | 19% | 12% | Falling |
| Personal use share (Claude.ai) | 35% | 42% | Rising |
| Coding share (Claude.ai) | — | 35% | Dominant |
| US top-5 states share | 30% (Aug 2025) | 24% | Deconcentrating |
| Global top-20 countries share | 45% | 48% | Slightly concentrating |
| Avg. task hourly wage | $49.30 | $47.90 | Slight decline |
| Avg. education required | 12.2 yrs | 11.9 yrs | Slight decline |
The report describes divergent patterns by surface. Claude.ai is diversifying and lightly consumerizing, with more personal use, a broader task spread, and slightly lower-skill tasks on average. The API is concentrating, with fewer task types accounting for more of the traffic. Anthropic attributes the decline in the Claude.ai skill mix partly to coding workloads migrating from Claude.ai to API-based agent tooling such as Claude Code.
Coding is 35% of Claude.ai conversations, the single largest category. Anthropic presents the combination of coding dominance, growing API concentration, and the coding-migration observation as first-party evidence that agentic coding is the principal deployment pathway for Claude (see Agentic AI).
Geographic distribution
US per-capita usage deconcentrated over the period: the top-5 states fell from 30% (August 2025) to 24% (February 2026). The Gini coefficient continued falling, but the report extends its projected time-to-parity from the January report's 2–5 years to 5–9 years, because the convergence rate slowed. The direction matches the January 2026 report; the pace is slower.
The global pattern runs opposite to the US. The top-20 countries rose from 45% to 48% of usage, a slight international re-concentration alongside US deconcentration.
Learning-curve findings
The report's central contribution is an analysis of user tenure. Compared with new users, users with 6+ months of tenure show:
| Dimension | High-tenure users |
|---|---|
| Success rate | +10% |
| Work share | +7 pp |
| Personal share | −10% |
| Education level in inputs | +6% |
| Top-10 task concentration | 20.7% (vs 22.2% newer) |
| Interaction mode | More collaborative, less directive |
| Model choice | Higher Opus usage on high-value tasks |
For causal identification, Anthropic runs regression controls for task (O*NET category and request cluster), model selection, country, language, and use case. Tenured users still hold 3–4 pp higher success rates after these controls, which Anthropic interprets as learning-by-doing rather than selection. A secondary education finding is that the years of education required by prompts rises about 1 year per additional year of Claude use; the newest users are 44% personal and one-year-tenure users 38% personal.
On model selection, a +$10 increase in hourly task wage is associated with +1.5 pp Opus usage on Claude.ai and +2.8 pp on the API. API users are about twice as wage-sensitive in model choice as Claude.ai users, which the report describes as consistent with production workloads optimizing unit economics.
Emerging API workflows
Two API workflow categories doubled since November 2025: business sales and outreach (lead qualification, data enrichment, cold-email drafting) and automated trading and market operations (monitoring, proposal generation, trader notifications). The report characterizes both as high-autonomy, low-human-in-the-loop use cases, reinforcing a split between API automation and Claude.ai augmentation.
Key claims and confidence
| Claim | Confidence | Notes |
|---|---|---|
| Top-10 Claude.ai task share fell 24% → 19% Nov 2025 → Feb 2026 | high | Consistent classifier; same methodology as Jan report |
| Coding is 35% of Claude.ai conversations | high | Direct measurement |
| US top-5 state share fell 30% → 24% | high | Cross-report consistent direction |
| 6+ month users show 3–4 pp higher success after controls | medium–high | Identification via regression controls; not an RCT, but a meaningful within-population comparison |
| Learning-by-doing, not selection, drives tenure effects | medium | Anthropic's interpretation; plausible but selection channels cannot be fully ruled out |
| Global top-20 concentration rose 45% → 48% | high | Direct measurement |
| Convergence-to-parity timeline now 5–9 yrs (was 2–5) | medium | Very short observation window; high sensitivity to period endpoints |
Reception and reading against other sources
As with the January report, the first-party measurement caveats apply: the classifier is Anthropic's and the denominator is Anthropic's, which carries the same self-serving framing risks. The learning-curve claim survives controls for task, model, country, language, and use case, which makes it harder to attribute to product marketing than a headline adoption statistic; it is also less flattering than a simple "success rate is high" claim, in that it implies most users are not yet getting the full value.
The tenure-controlled success-rate analysis is a methodologically different evidence class from capability benchmarks such as GDPval and METR horizons (see AI Benchmarks and Evaluation), because it measures user-plus-model performance in the wild rather than model capability in isolation.
The report bears on several wiki pages. For AI and Productivity, it adds a learning-curve channel the page previously lacked: Brynjolfsson's "AI as leveler" is primarily a same-day productivity comparison, whereas the Anthropic tenure analysis shows a within-user productivity slope over months, distinct from and complementary to Copilot studies, AMA sentiment, and NANDA enterprise-portfolio findings. For Agentic AI, coding at 35% of Claude.ai conversations is a first-party quantitative anchor for agentic deployment concentration, to be combined with API concentration growth and coding migration. For AI Labor Disruption, the coding share together with geographic and occupational deconcentration fits a pattern of rapid, broadening deployment rather than siloed early-adopter concentration.
On the Enterprise AI Deployment Gap, the doubling of emerging API workflows (sales enablement, automated trading) over three months is one data point against a "stalled enterprise AI" framing, but does not overturn the NANDA finding: these are API-side workloads often routed through vendor platforms, consistent with NANDA's own result that vendor-built paths succeed roughly twice as often as internal builds. NANDA's 95% enterprise-pilot-failure finding is not contradicted by Anthropic's API-workflow growth, because the two measure different denominators — NANDA measures organizations attempting internal builds, while Anthropic measures traffic across all paths, including the vendor-mediated ones NANDA finds succeed disproportionately.
The report's learning curve also sits in tension with Lancet deskilling. Anthropic's learning curve is with-AI performance improving over time; the Lancet finding is without-AI performance degrading. Both can hold simultaneously: prolonged use may improve combined performance while eroding the unaided baseline. The Economic Index does not measure unaided performance, so by design it cannot observe Lancet-style deskilling.
A terminological caution applies to the word "deskilling." The Anthropic Economic Index uses it in the January 2026 report's sense, where the AI performs the higher-education task within a job — a task-composition claim. That is distinct from the Lancet/AMA definition, which describes erosion of unaided human performance over time, and distinct again from the March report's learning-curve finding of with-AI performance improving. When comparing across sources, "deskilling" is used here in the Lancet/AMA sense, and occurrences in the Anthropic Index should be read in the task-composition sense unless context indicates otherwise. See AI Deskilling for the full disambiguation.
Read against forecasts, the report's adoption trajectory can be compared with Stanford HAI AI Index Report 2026, the grading of AI 2027's 2025 predictions (Source: blog.aifutures.org), and Situational Awareness: The Decade Ahead.
Relationships
- supports: AI and Productivity — adds learning-curve channel and deployment-composition trend data
- supports: Agentic AI — first-party anchor for coding-share concentration
- depends-on: Anthropic Economic Index — January 2026: Economic Primitives — methodological predecessor
- related: MIT NANDA — The GenAI Divide (State of AI in Business 2025), Enterprise AI Deployment Gap — complementary enterprise-portfolio view
- related: Lancet Endoscopist Deskilling Study (2025), AI Deskilling — contrast: with-AI learning vs. without-AI deskilling
- related: Stanford HAI AI Index Report 2026, (Source: blog.aifutures.org), Situational Awareness: The Decade Ahead — adoption trajectory relative to forecasts