Conversation Artifacts

Sources#

Summary#

The Chapter-2 contribution of the Anthropic Economic Index's June 2026 Cadences report: a new primitive — the artifact, the primary output a conversation produces (a document, an explanation, a piece of code, an academic paper). Where prior reports classified the task or request, this one classifies what the user walks away with, into 30+ categories. It makes AI's economic output legible, and surfaces two sharp regularities: compute scales with the value of the work, and Claude answers above the level it was asked.

Evidence note. empirical — privacy-preserving classifiers over chat and Cowork conversations sampled April 10–June 10, 2026; wages from BLS OEWS (May 2025); token counts use geometric means (right-skewed) and are not model-adjusted. First-party; classifier-inferred. See Anthropic Economic Index.

The artifact as unit of analysis#

93% of Claude conversations produce an artifact. The mix:

Artifact	Share	Family
Explanations	17%	conversational (~⅓ total)
Documents & reports	15%	written deliverable (~⅓ total)
Guidance	11%	conversational
Code / apps / scripts	—	technical (~⅙ total)

What an output is doesn't tell you what it's for. The same artifact can be a work deliverable or a personal project, so the report cross-cuts artifact type with the work/personal/coursework split from earlier reports:

Almost always personal (>80%): creative writing, guidance, recipes.
Mostly work (~80%): marketing content (80%), blogs/articles (81%), database queries (82%).
Split ~50/50: plans and strategies (44% work / 49% personal), translation (42% / 44%).

Flipping the question — what each use produces: work conversations most often yield documents/reports (20%); coursework yields documents (21%) and explanations (20%); personal rarely produces a document (6%), leaning instead to explanations (25%) and recommendations (22%).

Compute tracks the value of the work#

The report's most economically loaded finding: the tokens a conversation consumes rise with the estimated value of its output. Mapping each work conversation to the occupation that typically performs its task, median tokens rise with the occupation's median wage — marketing managers earn ~2× editors and their conversations use ~2.5× the tokens (noisy, with outliers — pharmacist-mapped conversations use far fewer tokens than the wage would predict). By artifact: building apps uses >3× the median conversation's tokens; a typical explanation uses ~⅕. About 44% of the wage–token gradient is explained by output mix — higher-wage occupations produce more compute-intensive artifacts.

Decomposing why higher-wage conversations cost more (Table 2.4, occupation wage terciles, bottom = 1×):

Measure	Bottom third	Middle	Top third
Tokens per conversation	1×	1.80×	2.07×
Turns per conversation	1×	1.58×	1.53×
Claude's response per turn	1×	1.25×	1.34×
Price-weighted compute cost	1×	1.98×	2.05×
Extended thinking enabled	31%	33%	34%

The report's reading is labor-augmenting, not labor-displacing: Claude produces more and users engage more (more turns) in high-value work — production from both sides moves together, so "the human remains involved in the highest-value tasks." Compute intensity and delegation also co-move: across artifacts, mean AI autonomy and median token use rise together (r = 0.68).

Replicated on a second lab's telemetry, in the one direction it was easy to check (September 2026). Google ATLAS's science report isolates ~360,000 scientific interactions out of its 15M-interaction Gemini corpus and compares them to the average work conversation on the same surfaces: +19% tokens, +11% turns, +7% multimodal share, and +26% on ATLAS's domain-expertise score (§3.1, conversational surfaces only). Different product, different classifier, different unit of "value" — a task-sophistication score and an occupational over-representation rather than an occupational wage — and the same co-movement of compute, turns and the apparent demandingness of the work. The limits of the corroboration are worth stating: ATLAS reports no wage cut on this slice, tokens are compared against a work baseline rather than across terciles, and the expertise score is a classifier judgment of the prompt, so this supports the "compute rises with the value of the work" direction without replicating the wage gradient itself.

Claude answers above the level it was asked#

A reading-level classifier estimates the years of education needed to understand the prompt and, separately, Claude's response. In almost every category Claude's output sits ~1 education-year above the prompt. The gap is widest where users describe something to be built — image & graphics (+2.6 years), games (+1.9), apps & websites (+1.7) — and near-zero for audience-facing writing (blogs −0.1, academic papers +0.0, email +0.3), where prompts already draft language in the target register. An academic-paper output needs 16+ years of education (bachelor's-plus); 15% are PhD-level (20+ years).

Why it matters#

The artifact lens is the output-side complement to the time-side cadences: two resolutions the report adds to make AI's economic footprint legible. The compute-tracks-value regularity is the one to watch — it says the price of an AI-produced output is beginning to encode its economic worth, a market-like signal emerging inside usage logs.

Connections#

The Household Production Boundary — where most artifacts are actually produced: 86.5% of conversational AI usage is non-work, and the meal plans, budgets, and filled forms it yields are exactly the value national accounts do not record
Anthropic Economic Index — the research program; this is its Chapter-2 contribution (the artifact classifier)
AI Usage Cadences — the sibling Chapter-1 advance: finer resolution on when, as this is finer resolution on what
AI Adoption in Scientific Work — the cross-lab check on this page's central regularity: on Google's telemetry, scientific interactions run 19% more tokens and 11% more turns than the average work conversation
Conversation-to-Delegation Shift — the autonomy/delegation lens; compute and delegation co-move (r = 0.68), and the report's autonomy-by-surface finding lives there
Returns to Expertise in Agentic Coding — "the human stays involved in high-value work" is the augmentation reading of the same expertise-amplifies-the-agent story
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — artifacts are the output side of what "AI can do a task" means; complements the occupational-exposure measures
The Automation–Optimism Link — the perceptions companion; heavy delegators produce more compute-intensive artifacts and feel more optimistic
AI as Primary Author — the artifact is authorship moving to the model; the reading-level lift (+1 year) is one measure of how far
Verification as the New Bottleneck — a legible artifact is what a human must review; classifying the output is a step toward instrumenting that check

Open Questions#

Tokens are a proxy for both compute cost and output value, but verbose models inflate tokens per unit of intent (the same critique Conversation-to-Delegation Shift raises); how much of "compute tracks value" is genuine value vs. models simply emitting more?
The reading-level "+1 year" gap may be register (terse prompts, polished replies) rather than substance; can it be separated from genuine elevation of content?
Artifact classification is first-party and single-model-graded; do the 30+ categories and the work/personal/coursework split survive independent replication?

Sources#

Anthropic Economic Index report: Cadences — Anthropic Economic Index report: Cadences (June 26, 2026), Chapter 2 "Artifacts": §What is each artifact used for, §Cost tracks the value of the work (Table 2.4), §Claude answers above the level it was asked
Google AI & Economy ATLAS: AI in Science (September 2026) — AI in Science: Early Insights (Google, Google DeepMind, MIT FutureTech, September 2026, 42pp; empirical with vendor COI). Cited here for §3.1 only: the token, turn, multimodality and domain-expertise deltas of science interactions against the average work conversation. Full treatment on AI Adoption in Scientific Work