Anthropic Economic Index report: Cadences (June 2026)

Anthropic Economic Index, June 26 2026 issue. The third in the EI series, and the first to (a) sample at a high enough rate to study hourly patterns, (b) classify the artifact each conversation produces, (c) break out Claude Code and Claude Cowork separately from chat + 1P API, and (d) publish first results from the EI Survey linking Claude-usage data to self-reported expectations about AI and work.

The framing line from the introduction sets the scope: "With the rapid growth of Claude Code and Cowork, Claude sessions now increasingly consist of long-running agentic tasks. Chat transcripts no longer fully capture how people are using AI, and our methods for studying Claude's economic impacts have had to adapt." The whole report is Anthropic catching up with their own tasks → responsibilities product surface.

What's new methodologically

Higher sample rate → hourly resolution (prior EI reports used 7-day samples)
New classifier labels the artifact each conversation produces (>30 artifact categories — see Artifacts (Claude Output))
Three-way data split: Claude conversations (chat + Cowork) / Claude Code / 1P API, aggregated monthly
Plus the EI Survey (launched April 2026), linking Claude-usage to self-reported expectations via Anthropic's privacy-preserving system

Chapter 1 — Cadences: how the rhythms of life shape Claude usage

See AI Usage Cadences for the named concept. Key findings:

Pattern	Datapoint
Workweek vs weekend (personal-use share)	~35% weekday → just under 50% weekend; biggest swing in high-income countries
Weekend Claude Code clusters that fall	backend architecture, API debugging, data storage
Weekend Claude Code clusters that rise	AI agent design, quant trading, gaming; starting-a-business queries peak Sat/Sun globally
Hourly peaks	News 7am · business correspondence 10–11am · recipes 2.3× at 6pm · media recs evening · sleep advice in pre-dawn hours
Night/weekend work skew	Tasks skew to higher-wage occupations (marketing managers, programmers); bottom two wage quartiles fall as share — robust to removing tech occupations
Tax-day spike	Tax-related conversations 8× the May average on April 14, persisted April 15, dropped sharply April 16

Chapter 2 — Artifacts: what Claude actually produces

See Artifacts (Claude Output) for the named concept.

93% of Claude conversations produce an artifact that the classifier can identify
Top categories: explanations (17%) · documents and reports (15%) · guidance (11%)
Two big buckets ≈ ⅔ of all artifacts: conversational outputs (explanations / guidance) and written deliverables (documents / presentations). Code + technical artifacts ≈ ⅙.

Use-case split per artifact (work / personal / coursework classifier applied):

Almost always personal: creative writing, guidance, recipes (each >80% personal)
Almost always work: marketing content (80%), blogs/articles (81%), database queries (82%)
Mixed: plans/strategies (44% work / 49% personal), translations (42% / 44%)
Coursework signature: academic papers, theses, educational materials, math queries

Flipping the lens — what each use case tends to produce:

Work → documents/reports (20%), explanations (9%), email drafts (7%), analyses (6%)
Coursework → docs/reports (21%), explanations (20%), educational materials (11%), papers (6%)
Personal → explanations (25%), recommendations (22%), docs only 6%

Cost tracks the value of work

Median tokens-per-conversation, mapped to the median wage of the occupation that typically performs the underlying task, traces a positive relationship:

Marketing managers earn ~2× what editors do ($80 vs $37/hour) → conversations use ~2.5× the tokens
Pharmacists ~3× statistical assistants ($68 vs $24/hour) → conversations also use more tokens (source body truncated mid-quantification)

This is the first empirical anchor in this vault for compute scaling with the value of the work being done — see the update on Code Is Free. It's the second-half completion of Lopopolo's "code is free": cost moves to where the value is.

Chapter 3 — EI Survey: expectations and experience by usage cluster

The first results from the Anthropic Economic Index Survey (launched April 2026), linked through the privacy-preserving system to actual usage data. The headline finding the report previews up front:

"Expectations and experiences vary systematically with how people use Claude: people who use Claude in the most automated way expect AI to take on more of their tasks in the next year, yet feel the most optimistic about what that means for their work, anticipating positive impacts on pay, job security, and meaning."

This is a non-obvious correlation worth holding onto — see FOBO (Fear of Becoming Obsolete) for why. The lazy reading is "people delegating most to AI are walking into their own redundancy"; the survey-linked finding is the opposite. Either selection (optimists already delegate more) or experience (delegation reveals delegation is good for me) — the report previews the finding but the body is truncated before the mechanism is laid out.

warning Source truncated The captured body stops mid-Chapter-2 ("pharmacists earn nearly three times…"). Chapter 3 findings are reconstructed from Anthropic's announced preview at the top of the report. Mechanism, sample size, and cluster definitions for the EI Survey need a second pass from the full PDF if cited externally.

Why this matters for this vault

Claude Cowork is now first-class. Anthropic is splitting it out from chat in their own reports. The product surface this vault has been tracking via Sandeep's 5-surface stack and Cvent's CISO refusal is now appearing in Anthropic's own usage data as a distinct cohort.
Artifacts is the new unit Anthropic uses internally. Not "tasks", not "prompts" — the output an LLM produces in a session. This may shape how the next generation of product analytics and pricing work. Worth watching whether competitors adopt the framing.
The compute-vs-wage scaling is the missing empirical anchor for Code Is Free. Lopopolo / Cherny / Karpathy said code is free; Anthropic now has the matched-distribution data showing that the cost of producing AI output moves with the value of the work being done. Implication: budget by occupation, not by team.
The automation-equals-optimism survey result is the empirical counter-narrative to the FOBO literature. Useful for IT-LT conversations where the FOBO-driven anxiety reading dominates the discussion (Designing IT Roles for an AI Era (Talent Strategy POV)). Carries a flagged-confidence asterisk: linked data, not external causal claim.

Connects to your work

This is the closest external research lands to the user's day job:

Cowork separated from chat in usage data = the IT exec's "what should we let employees do with Claude" question now has Anthropic's own break-out.
The artifact taxonomy is a clean way to size AI usage at P&G: not by seats or by tokens, but by artifact mix. Imagine the same breakdown over Manila IT's six months of Cowork traffic.
The automation → optimism correlation gives the IT-LT a piece of empirical air to push back on the dominant headline FOBO narrative — without dismissing the underlying anxiety the FOBO (Fear of Becoming Obsolete) page maintains.

Sources

Anthropic Economic Index report: Cadences — June 26, 2026
Telegram capture #3480 — user-shared via Telegram for ingest