The Claude-usage landscape
Clustering US states and countries by how they use Claude — UMAP + HDBSCAN on Anthropic's Economic Index open data (April–May 2026). The method is described at the bottom of the page.
Hover a cluster to see its details; click a dot (or search on the left) to see the state itself.
A note on Hawaii
On this map it sits among the “Study and analysis” states, yet it belongs to “Business first” (cluster fit 1.0). The clustering was done in a much higher-dimensional space, and flattening to two dimensions cannot preserve every distance — imagine folding the chart so Hawaii lands near the other “Business first” states. The placement still says something real: among “Business first” states, Hawaii is the one that most resembles “Study and analysis.”
Hover a cluster to see its details; click a dot (or search on the left) to see the country itself. Zoom in to see full country names.
Method
Each geography is described by ~90 features from Anthropic's Economic Index release (collaboration patterns, automation vs. augmentation, artifact types, use-case splits, and its occupation / request-topic / work-activity mixes — restricted to categories published for at least 90% of geographies, so data coverage does not masquerade as usage style). Features are z-scored and block-weighted, embedded with UMAP (cosine metric), and clustered with HDBSCAN; hyperparameters were selected by Bayesian optimization (Optuna TPE, 150 trials per level) against a single objective combining density-based cluster validity (DBCV), the fraction of unclustered points, and stability across UMAP random seeds. The map shows a 2-D UMAP view of the same features. Gray points are HDBSCAN outliers that fit no cluster. Usage volume was excluded from the features; clusters describe how, not how much. Countries first split coarsely into a developed/developing pair; the granularity shown here was deliberately selected below that split. Hover for details; scroll to zoom; click a point to highlight its cluster, click empty space to reset.