A model maker's own usage index counts conversations, not the economy

The setup

On 10 February 2025, Anthropic published the first edition of the Anthropic Economic Index, a periodic report meant to track how its Claude models get used across occupations. The method, described in an accompanying working paper, sorts several million anonymized Claude.ai conversations into task categories drawn from O*NET, the U.S. Department of Labor's standard occupational database. Anthropic built an internal tool, Clio, to do this classification without a human researcher reading the original text. The company also released an aggregated version of the dataset for outside researchers to check its own work, a step the report frames as necessary given the obvious conflict in a company studying its own product.

What the documents show

The paper's own numbers are specific about what they cover: software development and writing tasks together account for close to half of the classified usage, about 36% of occupations show AI use on at least a quarter of their associated tasks, and just under three in five conversations look more like augmentation than full automation of a task. Every one of these figures is Anthropic's own measurement of Anthropic's own product traffic, not an independent audit, though the released dataset lets other researchers rerun the classification themselves.

The friction

The paper is explicit about what it cannot see. It cannot tell whether a person used Claude for paid work or a side project, whether an automation-coded reply was pasted unedited into a deliverable, or how representative Claude's user base is of AI use generally, given that Anthropic markets Claude heavily to developers. Its authors describe the findings as painting a picture of usage on a single platform, language that rules out extending the results to ChatGPT, Gemini, Copilot or any other product's users.

What changed in the work

What the index adds to the record is a task-level view that a subscriber survey cannot easily produce: instead of asking someone what they think they use a chatbot for, it classifies what conversations already say. That is a narrower but more concrete unit of measurement than a self-reported adoption percentage. Reading it as a national labor-market indicator, rather than a usage log for one company's product, is an extension the documents themselves do not support, and is an editorial line worth holding when the report gets cited elsewhere.

• Does a cited AI-usage figure come from one vendor's logs, a survey of opinions, or a government data series?
• Would the occupational categories used here match how a reader's own job actually gets described in a similar dataset?
• What does the source say it cannot measure, and has that limit survived being repeated into a headline?

Anthropic says it plans to keep publishing the index and expanding the released dataset, which at least gives outside researchers a way to test claims rather than take them on faith. Until other model providers publish comparably detailed, comparably transparent breakdowns of their own traffic, the index remains a well-documented single-company case study rather than a market-wide measurement.