Anthropic is now measuring itself by the hour.
The company’s June 26 Economic Index report, titled “Cadences,” introduces a data pipeline that samples Claude usage at hourly resolution, a methodological shift from prior aggregate reporting. The report also introduces a new classifier that labels the output of each conversation. These aren’t aesthetic changes. Hourly granularity and output-level classification are what you’d build if you wanted to track agentic sessions, which run longer, produce more output, and behave differently from a conversational exchange.
According to Anthropic’s report, Claude sessions increasingly consist of long-running agentic tasks. The primary drivers are Claude Code and Cowork growth. The company also breaks out data separately for Claude chat and Cowork conversations versus the 1P API at monthly aggregate level. These are vendor-reported findings. Anthropic is reporting on its own platform usage; no independent researcher has validated the underlying dataset or methodology.
Why it matters for markets: the chat-to-agentic transition has been described qualitatively across multiple recent cycles. Analysts have cited agentic AI adoption in enterprise software, infrastructure spend, and pricing model shifts. Cadences is the first primary-source document from a frontier lab that attempts to quantify this shift at session level, using hourly sampling. It’s not independent data. But it’s the most granular vendor disclosure on usage composition published to date, and it establishes a baseline for future comparison.
The real story is the methodology upgrade. Hourly sampling plus output-level classification means Anthropic can now distinguish between a 90-second chat query and a 4-hour coding session that generates 50,000 tokens. Prior reporting couldn’t make that distinction reliably. If the pattern the company is describing is real, this is the tooling that would expose it, and that makes Cadences a useful document even accounting for its self-reported limitations.
What to Watch
What to watch
whether independent researchers publish analyses of agentic AI usage patterns that corroborate or contest Anthropic’s directional findings. The Economic Index has been published across prior cycles; tracking how the methodology and findings evolve over subsequent reports will matter more than any single release. Enterprise AI buyers should treat Cadences as a signal about how Anthropic wants to position Claude for agentic workloads, useful context for pricing and integration planning, not a substitute for independent assessment.
The catch is the verification gap. Anthropic’s data describes Anthropic’s platform. Claude Code growth drives agentic session counts on Claude. That’s not a reason to dismiss the findings, it’s a reason to hold them at the appropriate confidence level and watch for corroboration from sources with no stake in the conclusion.