Summary: On March 24, Anthropic published Learning curves, an Economic Index study of Claude usage across February 2026. The article says the report is most useful as part of a comparable longitudinal series rather than as a single-month snapshot. It also notes that Opus 4.6 arrived on February 5 and Sonnet 4.6 became the default on February 17, so product changes could affect the month’s patterns.
The Economic Index series picked up a new entry on March 24, when Anthropic released “Learning curves,” a study of how people used Claude across all of February 2026.
One trait has defined this series from its very first report, and it defines the new one too: the raw material is neither polls nor interview transcripts but conversations that genuinely took place on Anthropic’s platform. The fourth edition, which came out in January, had only just added the economic primitives classification, placing 56% of conversations under augmentation and 41% under automation. “Learning curves” stays on that track, with February now under the lens.
Evidence Built from What People Do
A poll records how people describe their AI habits; usage logs show what those habits actually look like. That difference is exactly where the series earns its keep: adoption, properly understood, is not a number in an intentions poll but an action performed day after day.
The caveats belong in the same reading. Claude’s users lean toward technical and knowledge work, so the conclusions do not transfer automatically to every industry, and a sample drawn from a single vendor bears the imprint of one product’s design choices. Held onto as a longitudinal record of AI’s impact on work, it repays the attention far better than any lone cross-section. The series’ steadiness also compounds: the classification added in January can be run again on every new month, so each report extends a genuinely comparable time series — an uncommon thing in a field where methods tend to be reinvented every publication cycle.
Reading It with the Right Expectations
- Follow the direction across editions instead of one month in isolation: a single month’s distribution gets contaminated by events, and consecutive editions are where the signal lives
- Put the product calendar alongside it: when a usage shift lands relative to a feature launch is often more telling than the absolute numbers
- Keep an eye on the task mix: which work moves over to AI and which stays in human hands — how that boundary travels is the real story
As it happens, February was an eventful month to sample: Opus 4.6 landed on February 5, and Sonnet 4.6 took over on February 17 as the default model. The usage patterns of that month will all but certainly bear traces of both launches. Keep that context in view, and a product switch will not be mistaken for a genuine behavior change. The title earns its place here as well: learning curves describe capability rising with use, and the readings that matter most are organizational — which teams steepen their curve, and which plateau.
What Product and Hiring Teams Can Take Away
The real distribution of tasks is laid out for product teams in behavioral terms: which scenarios already absorb heavy delegation and which keep holding out — one more input for roadmap priorities, and one more honest than launch-day applause. Those planning hiring and training gain in equal measure: on which skill floors are being raised, usage data tells the truth more reliably than a manager’s intuition.
No single report hands you the answer. What it does is move “AI’s impact on work” out of the realm of impressions and into a series you can track — and in a 2026 crowded with concept-stock narratives, that alone puts it ahead of most industry surveys. For Anthropic itself, the series works as a form of public accountability: the same usage data that shows where Claude works shows where it does not, and putting it out on a steady cadence sets a bar the company has implicitly agreed to be measured against.