|
The median chat session that produces a blog post runs 13 rounds of back and forth. The median Claude Code session that produces the same artifact contains a single human prompt. Same lab, same model family, twelve fewer rounds of human judgment.
That number lives in the Anthropic Economic Index, the dataset Anthropic has published since early 2025 to measure how AI is actually used across the economy. As of July 22, you can query the whole thing in plain English from inside Claude.
|
Same artifact, two surfaces
|
Chat and Cowork
13
median rounds of
back and forth
|
⇄ |
Claude Code
1
median human
prompt
|
Median session producing a blog post or article
Autonomy gap, Claude Code over chat and Cowork (1 to 5 scale)
| Scripts and code snippets |
|
+0.53 |
| All output types, average |
|
+0.37 |
| Holding the model fixed (Sonnet) |
|
+0.26 |
Source: Anthropic Economic Index report, Cadences, June 26, 2026. Autonomy scored 1 to 5 by classifier.
|
A dataset became a tool call
Anthropic released the Economic Index connector on July 22, 2026. You enable it from the connectors menu in claude.ai, it works in any conversation with any Claude model, and there is nothing to install. Ask which occupations use AI the most, or what tasks teachers bring to Claude, and the answer comes from Index data instead of model recall.
The shape of this matters more than the feature. Claude connectors are servers speaking the Model Context Protocol, the open standard Anthropic published for wiring models to external data and tools. So a research dataset that used to reach people as a report plus a download is a queryable resource that any conversation can pull from mid-thought.
|
Index datasets
Reports, dashboards, downloads
|
→ |
Connector
Protocol server in the directory
|
→ |
Grounded answer
With a pointer to the source rows
|
One caveat rides along with every answer. The Index measures patterns in Claude usage, not the labor market as a whole, and Anthropic says Claude will point back to the source data and its limits as you explore. Read it as telemetry from one large deployment, not a census of work.
What is actually in the data
The most recent release is the June 26, 2026 report, titled Cadences. It covers chat and Cowork conversations sampled between April 10 and June 10, 2026, with Claude Code sessions and first-party API traffic broken out separately. Two changes in that release make it useful to builders rather than just to economists.
First, sampling runs continuously instead of in seven-day windows, which exposes hourly rhythms. Recipe requests are 2.3 times more frequent at 6 p.m. than average. Tax clusters ran eight times higher on April 14 than on a typical May day. Second, a new classifier labels the artifact each conversation produced, sorted into more than 30 categories.
|
93%
of conversations produced
an identifiable artifact
|
17%
explanations, the single
most common output
|
15%
documents and reports,
then guidance at 11%
|
Conversational outputs like explanations and guidance account for roughly a third of conversations. Written deliverables like documents and presentations account for another third. Code and technical work covers about a sixth. That is the mix in consumer chat and Cowork traffic, which is a useful prior if you are guessing what your own users will ask a model to make.
The surface moves delegation more than the model does
Anthropic scored every conversation for AI autonomy on a 1 to 5 scale, running from none to extreme. Low-autonomy work is easy to specify: math, translation, question answering. High-autonomy work requires selecting among many valid choices: apps, websites, games, presentations. Claude Code scored higher on 26 of the 31 output types measured.
The obvious explanation is model choice. Opus served 54% of Claude Code sessions against 10% of chat and Cowork conversations. Hold the model fixed and the gap survives anyway: among Sonnet-served conversations, Claude Code still shows 0.26 points more autonomy, which Anthropic reads as the product mattering more than the underlying model.
About two thirds of the total 0.37-point gap comes from identical tasks being executed with more delegation. The remaining third comes from a different mix of outputs across surfaces. The blog post case is the cleanest cut of it, since the underlying requests look similar on both surfaces and the working style does not.
|
Key insight: This is a measurement of harness quality, not checkpoint quality. How much context the agent gets up front, whether it can act without asking, and what it can verify on its own decide how much a person hands over.
|
Compute tracks the value of the output
Token consumption is wildly uneven across artifact types, and the spread is the part worth internalizing if you price a product. Conversations that build apps use more than three times the tokens of the median conversation. A typical explanation uses about a fifth.
|
Tokens per conversation, relative to the median
| Building apps |
|
3.0x + |
| Median conversation |
|
1.0x |
| Typical explanation |
|
0.2x |
|
2.07x
tokens used by top wage tercile occupations versus bottom tercile
|
44%
of that wage gradient is explained by output mix alone
|
Source: Anthropic Economic Index report, Cadences. Work-related chat and Cowork conversations, geometric means.
|
Autonomy and token use move together, correlating at 0.68 across artifact types on chat and Cowork. The artifacts people delegate hardest are the same ones that burn the most compute, which is intuitive and also inconvenient: the sessions with the thinnest human supervision are the expensive ones.
Higher-wage conversations are not simply longer prompts either. Claude produces 1.34 times as much output per turn, users take 1.53 times as many turns, and extended thinking is switched on in 34% of those conversations versus 31%. Both sides do more work at once, which Anthropic reads as a pattern that augments rather than displaces.
Claude answers above the level it was asked
A classifier estimated the reading level of each prompt and each response, expressed as years of education needed to follow the text. In nearly every category, the response sits about one year above the prompt. Prompt and response levels correlate at 0.87, so the two rise together.
The gap widens where a user describes something to be built: images and graphics at plus 2.6 years, games at plus 1.9, apps and websites at plus 1.7. It sits near zero for audience-facing writing, with blogs at minus 0.1 and academic papers at plus 0.0. Part of that is register, since prompts are terse and responses are polished. The rest is a decent proxy for how underspecified build requests tend to be.
What the heaviest delegators say
The third chapter links roughly 9,700 survey responses to real usage data through Anthropic's privacy-preserving analysis system. The sample is not representative of the workforce: computer and mathematical occupations make up about 30% of respondents against 4% of US employment. Read it as a picture of technical early adopters.
Close to 6 in 10 respondents expect AI to handle a larger share of their work tasks 12 months from today, and over a third expect it to handle most or nearly all of them. 10% rated losing a job they wanted to keep in the next year as likely or very likely. Respondents worried more about junior colleagues than about themselves.
Then the counterintuitive result. Across all six dimensions measured (pay, job security, ability to find a new job, meaning, autonomy, and human interaction), respondents with a higher share of automated sessions reported more positive expectations. The relationship holds when controlling for how long someone has used Claude, which is the closest available proxy for enthusiasm.
Self-reported gains ran 86% for speed, 82% for scope, and 69% for quality. 68% report learning more with AI and 57% report that AI made their skills more valuable. Heavy delegators report learning at the same rate as everyone else, though Anthropic notes that self-assessment cannot rule out skill erosion.
Three things to take from this
|
1. Fix the loop before the model. One product change, same weights, 0.26 points of measured autonomy. If your agent is stuck in a 13-round conversation with its user, the scaffolding is the more likely culprit than the checkpoint.
|
|
2. The artifact is the cost unit. Apps burn more than three times the median conversation, explanations about a fifth, and 44% of the wage-token gradient is output mix. Model your unit economics per artifact produced, not per seat or per user.
|
|
3. Publishing data as a connector is a distribution choice. The same numbers that lived in a report and a download are a protocol server, answered at query time with a pointer back to source. Any team sitting on a dataset can copy the pattern, and the grounding beats hoping a model memorized the numbers.
|
A workflow that gets real answers
Anthropic suggests starting broad, then drilling in, then asking for the underlying data. That third step is the one people skip, and it is the one that separates a grounded answer from a plausible paragraph.
1. What does the Index say about my occupation
2. Break that down by task and by automation share
3. Show me the underlying rows and the sample window
The full datasets remain available for download on the Anthropic site, so anything the connector tells you can be checked against the release it came from.
The Index watches one deployment, but it is the most granular public record of what people hand to a model and what they keep for themselves. That line is moving one product surface at a time, not one model release at a time.
|
ResearchAudio.io
Sources: Anthropic, The Anthropic Economic Index connector (July 22, 2026), and Anthropic Economic Index report: Cadences by Massenkoff, Lyubich, Sacher, Hitzig, Zhang, Heller, and McCrory (June 26, 2026).
|
|
