Anthropic Economic Index Report: The Learning Curve

The Anthropic Economic Index's February 2026 sample analyzed 1 million conversations collected from Claude.ai and the 1P API, covering the period February 5-12. The sampling point comes 3 months after the Claude Opus 4.5 release and coincides with the Claude Opus 4.6 release. Overall, usage broadened, and task distribution became less concentrated especially on Claude.ai.

On Claude.ai, the share of the top 10 O*NET tasks fell from 24% to 19%. Coding shifted more toward the API, coursework declined from 19% to 12%, while personal use rose from 35% to 42%. As a result, the average economic value of tasks performed on Claude.ai dipped slightly from $49.3 to $47.9.

This shift reflects an influx of lower-wage personal queries and more diverse use cases. Average years of education fell from 12.2 years to 11.9 years, the autonomy users granted to AI increased, and the time a human would need to handle the task alone decreased by about 2 minutes. Conversely, tasks performed by Claude were rated slightly more difficult for humans to do without AI.

On Claude.ai, augmentation increased slightly, driven by growth in validation and learning patterns. On the 1P API, by contrast, automation decreased significantly.

In terms of the learning curve, experience improves outcomes. Users with 6+ months of tenure had 10% fewer personal conversations, 6% higher education level reflected in their inputs, and a 10% higher conversation success rate. This gap is not explained by task selection or country of origin, suggesting the possibility of genuine learning-by-doing as people become better users through repeated use.

Model choice is also tailored to the task. Paid Claude.ai users chose Opus for coding 4 percentage points more than average, and 7 percentage points less for tutoring-related tasks. On the API, this gap was about 2x more pronounced.

Geographic disparities remain large, but convergence continued within the US. Globally, the top 20 countries accounted for 48% of per-capita usage, up from 45% previously, while within the US, the share of the top 10 states fell from 40% to 38%.

On the API side, automation patterns became clearer. Two workflows at least doubled in share over the past 3 months:

• Business sales & outreach automation: sales enablement, B2B lead qualification, customer data enrichment, cold-email drafting
• Automated trading & market ops: monitoring markets/positions, investment proposals, communicating market conditions to traders