On March 24, Anthropic published the newest installment of its Economic Index series, titled “Learning curves,” analyzing Claude usage patterns across the full month of February 2026.
What has set this series apart since its first report still holds: the data is not surveys or interviews but actual conversations on Anthropic’s platform. The previous edition, published in January as the fourth in the series, had just introduced the economic primitives classification — 56% of conversations categorized as augmentation, 41% as automation. “Learning curves” continues that line with February in the frame.
A Report Written from Behavior Data
Surveys measure how people say they use AI; usage data measures how they actually use it. The gap between the two is precisely where this series earns its keep: adoption is not a percentage in an intent poll, it is behavior repeated daily.
The methodological limits belong in the same paragraph. Claude’s user base skews toward technical and knowledge work, so findings do not extrapolate cleanly to every industry, and a single-vendor sample carries the fingerprint of one product’s design. Tracking it as a longitudinal series on AI’s impact on work is far more useful than quoting any single cross-section. The series’ consistency is its compounding asset: because the classification introduced in January can be reapplied to each new month, the reports build a genuinely comparable time series — rare in a field where most studies change method with every publication cycle.
How to Read It Without Overreading
- Watch trends, not single points: one month’s distribution gets contaminated by events; the signal lives across consecutive editions
- Line it up against the product calendar: the timing relationship between usage shifts and feature launches often says more than absolute numbers
- Watch the task mix: which work gets handed to AI and which stays human — the movement of that boundary is the real story
February happens to be a lively sample month: Opus 4.6 arrived on February 5, and Sonnet 4.6 followed on February 17, becoming the default model. That month’s usage patterns almost inevitably carry the imprint of both launches. Read with that background in view, and you avoid mistaking a product switch for a behavior change. It is also why the report earns its title: learning curves describe capability rising with use, and the interesting readings here are organizational — which teams steepen their curve, and which plateau.
Value for Product and Hiring Decisions
Product teams get a behavioral view of real task distribution: which scenarios are already heavily delegated, which still resist — one more piece of evidence for roadmap priorities, and a more honest one than launch-day applause. Hiring and training planners benefit the same way: usage data is more truthful than managerial intuition about which skill floors are being raised.
One report will not hand you answers. But it turns “AI’s impact on work” from an impression into a trackable series — which, in a 2026 crowded with concept-stock narratives, already beats most industry surveys. For Anthropic itself, the series doubles as a public accountability mechanism: the usage data that shows where Claude works also shows where it does not, and publishing it on a regular cadence sets a bar the company implicitly agrees to be measured against.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.