Anthropic Financial Index report: Cadences

Introduction

One 12 months in the past, most Claude utilization took the type of a dialog between a consumer and an assistant. With the fast development of Claude Code and Cowork, Claude periods now more and more include long-running agentic duties. Chat transcripts not totally seize how individuals are utilizing AI, and our strategies for learning Claude’s financial impacts have needed to adapt.

To maintain tempo, we made a number of adjustments to our knowledge pipeline for the Financial Index. On this model, we:

• Pattern knowledge at the next fee, permitting us to view utilization patterns all the way down to the hourly degree.
• Introduce a brand new classifier that labels the output of every dialog.
• Share extra granular knowledge, breaking out outcomes for chat and Cowork conversations (collectively, “Claude conversations”) and the 1P API, aggregated at a month-to-month degree.¹

We describe further methodological adjustments within the Appendix. Collectively, these adjustments present a clearer image of how AI mirrors and diffuses into financial life.

As well as, we’ve beforehand lacked visibility into Claude’s affect exterior of consumer periods. How do individuals understand AI to be altering their work, or the alternatives out there to them? Does their utilization of AI form their expectations? In a perfect world, what would they need from AI? We report preliminary findings from the Anthropic Economic Index Survey, launched in April 2026.

We preview our most important findings beneath.

• In Chapter 1, we present how the rhythms of the exterior world form Claude utilization. Work-related queries subside on the weekend, although much less dramatically in probably the most extremely paid occupations; individuals are likely to ask for the information within the morning, and sleep recommendation peaks round 5 a.m.; tax-related requests surge round submitting deadlines.

• Chapter 2 explores the concrete outputs that folks take away from their Claude periods. These are extremely depending on what product they’re utilizing. Chat and Cowork present extra explanations than Claude Code, for instance. The character of the output additionally shapes individuals’s interactions with Claude. Constructing a web site leaves rather more to Claude’s judgment than translating a doc, the place the reply is basically decided by the textual content. We additionally see that extra compute is related to extra invaluable artifacts; the tokens a given output consumes rise with the estimated worth of the work.

• Chapter 3 presents the primary outcomes from the Anthropic Financial Index Survey, which we hyperlink to Claude utilization knowledge by means of our privacy-preserving system. Expectations and experiences range systematically with how individuals use Claude: individuals who use Claude in probably the most automated manner anticipate AI to tackle extra of their duties within the subsequent 12 months, but really feel probably the most optimistic about what which means for his or her work, anticipating optimistic impacts on pay, job safety, and which means.

Cadences

Our new privacy-preserving telemetry, which constantly samples a slice of conversations daily, permits us to review each day and hourly patterns in utilization, in distinction to the seven-day samples every earlier Financial Index report drew on. These analyses seize ebbs and flows in work patterns all over the world.²

We discover that Claude utilization mirrors the workweek, with private prompts spiking on the weekend. The hourly knowledge captures within-day patterns—individuals most frequently ask for sleep recommendation round 5 a.m. and for recipes round 6 p.m. We additionally see utilization reflecting key dates. For example, tax-related requests surged simply earlier than the US submitting deadline on April 15.

The workweek

The share of chat and Cowork³ conversations categorized as private use spikes from round 35% on weekdays to simply underneath 50% on weekends throughout the pattern interval (Determine 1.1). Exterior the workweek, customers’ conversations shift from enterprise correspondence, advertising and marketing copy, and slide decks to emotional help, medical questions, and funding recommendation. This shift is greatest for high-income international locations.

An identical sample is current in Claude Code and the 1P API visitors (i.e., API visitors routed straight by means of Anthropic), although each have decrease baseline charges of non-public use.⁴

Request clusters⁵ enable us to go one degree deeper and see which particular Claude Code duties swing most between weekdays and weekends. On weekends, the Claude Code utilization clusters that fall probably the most embody backend structure, API debugging, and knowledge storage. People who enhance probably the most embody AI agent design, quant buying and selling, and gaming.

Weekends can also create house for individuals to pursue new ventures. Throughout international locations, conversations associated to beginning a enterprise are highest on Saturday and Sunday. Nevertheless, job utility actions drop on the weekend together with different work-related duties.⁶

Every day rhythms

Hour by hour, Claude utilization displays the rhythms of each day life. Determine 1.2 exhibits the hourly frequency of various request clusters relative to their total common in world visitors.⁷

Folks ask for information at 7 a.m. native time. Enterprise correspondence (e.g., e-mail drafting) traces the arc of the workday, with a slight peak at 10–11 a.m. One of many greatest spikes is recipe requests, that are 2.3 instances extra frequent at 6 p.m. in comparison with the typical. Media suggestions are most concentrated within the night, whereas individuals search sleep recommendation within the few hours simply earlier than daybreak.

On nights and weekends, when individuals do flip to Claude for work, the duties skew towards higher-wage occupations (Determine 1.3). Whereas we won’t conclusively determine the roles of the individuals making these requests, this might replicate the truth that individuals in higher-paying occupations—like advertising and marketing managers or laptop programmers—usually tend to work exterior conventional hours. In distinction, duties associated to jobs within the backside two quartiles—like telemarketing and clerical work—fall to a smaller share of whole conversations. This sample is not pushed completely by laptop and mathematical duties: once we eliminated these occupations from the evaluation in a robustness verify, higher-quartile duties nonetheless elevated on nights and weekends.

Tax day

The pattern interval for this report covers tax submitting deadlines for individuals in america. Determine 1.4 exhibits a big spike within the share of tax-related conversations across the deadline. On April 14, tax-related clusters have been eight instances as widespread as on the typical day in Might and remained about as excessive on April 15. On April 16, they dropped sharply.

Artifacts

On this chapter, we classify every dialog on chat and Cowork (hereafter “Claude conversations”)⁸ by its artifact, which we type into greater than 30 classes. We confer with the first output Claude produces in a dialog—a doc, an evidence, a chunk of code, an instructional paper, and so forth, whether or not introduced in a chat window or as a separate doc—as an artifact. The complete listing of artifacts is within the Appendix.

Our classifier recognized 93% of Claude conversations as producing an artifact (Determine 2.1).⁹ The most typical artifacts are explanations (17% of conversations), paperwork and experiences (15%), and steerage (11%). Conversational outputs (like explanations or steerage) and written deliverables (like paperwork or displays) every account for a couple of third of conversations; code and technical work (like apps or scripts) for a couple of sixth.

What an output is does not let you know what it is for: the identical artifact may very well be a piece deliverable or a private mission. We have a look at that break up subsequent.

What’s every artifact used for?

Our January Economic Index launched a primitive that classifies every dialog as work, private, or coursework. Right here, we apply that break up to the artifacts produced in Claude conversations (Determine 2.2).

Some classes of artifacts are nearly all the time private. Greater than 80% of conversations producing artistic writing, steerage, and recipes have been categorised as private. Inside classes, the private and work-related makes use of can look fairly completely different. Private artistic writing, as an illustration, is dominated by fanfiction, worldbuilding, and poetry; the 13% that’s work-related is usually within the type of short-form video scripts, screenwriting, and speeches. Classes almost certainly to be work-related embody creating advertising and marketing content material (80%), creating blogs or articles (81%), and writing database queries (82%).

Many outputs are equally probably for use for private and work causes, together with creating plans or methods (44% work-related, 49% private) or translation (42% work, 44% private). For instance, the most typical sorts of private planning artifacts embody journey itineraries and exercise schedules, whereas work-related plans most frequently pertain to entrepreneurial or content material methods.

Lastly, artifacts which might be attribute of coursework embody creating tutorial papers and theses, instructional supplies and math-related queries, although a non-negligible share of every falls into each work and private classes.

We are able to additionally flip the query. As a substitute of asking what every output is used for, we will ask what kind of artifacts work, private, and coursework conversations every have a tendency to supply. Work conversations most frequently produce paperwork and experiences (20%), adopted by explanations (9%), e-mail drafts (7%), and analyses and summaries (6%). Coursework conversations look broadly related, with paperwork and experiences main there too (21%), intently adopted by explanations (20%), instructional supplies (11%), and tutorial papers (6%). In distinction—and unsurprisingly—solely 6% of non-public conversations produce a doc. As a substitute, the most typical outcomes are explanations (25%) and proposals (22%).

Value tracks the worth of labor

Producing these outputs requires compute, and we discover that compute tends to scale with the worth of the work. We measure every dialog’s computational prices in tokens—the quantity of textual content processed and generated, together with Claude’s inside reasoning—and examine throughout occupations by mapping every dialog’s categorised job to the occupation that sometimes performs it. All through this part, we prohibit our evaluation to work-related conversations.

The left panel of Determine 2.3 exhibits a optimistic relationship between the median conversation-level variety of tokens and the median wage in mapped occupation.¹⁰ For instance, advertising and marketing managers earn roughly twice as a lot as editors ($80 vs. $37 per hour) and conversations mapping to their duties devour roughly 2.5 instances as many tokens. Admittedly, the connection is noisy and there are notable outliers. Pharmacists, for instance, earn practically thrice what statistical assistants do ($68 vs. $24 per hour), but conversations mapped to pharmacist duties use solely about one twentieth as many tokens.

The tokens consumed to generate several types of artifacts inform an identical story. Extra difficult and invaluable outputs are likely to devour considerably extra tokens than less complicated outputs. For instance, conversations about constructing apps use greater than thrice the tokens of the median dialog. On the opposite finish of the spectrum, a typical clarification makes use of a couple of fifth of the tokens of the median dialog. About 44% of the wage gradient in token consumption is defined by output combine—larger wage occupations usually tend to produce compute-intensive artifacts.

Why does this matter economically? In conversations mapped to higher-wage occupations, Claude produces extra (1.34 instances as a lot output per flip), whereas customers have interaction extra (1.53 instances as many turns) and allow prolonged pondering extra often (34% of conversations versus 31%; Desk 2.4). Crucially, these transfer collectively: extra manufacturing from Claude doesn’t imply much less from the consumer. If the human stays concerned within the highest-value duties, the sample seems extra labor-augmenting than labor-displacing. It additionally exhibits that, to some extent, extra invaluable outputs price extra. The subsequent part examines how a lot of the decision-making inside every dialog is delegated to Claude.

How a lot autonomy does Claude need to determine by itself?

We measure this on a 1-5 scale, from “none” to “excessive.” Duties which might be simple to explain or specify contain little autonomy: the lowest-autonomy outputs are math or calculations, translations, and Q&As. Excessive-autonomy duties are people who require choice amongst many choices, e.g., creating apps and web sites, video games, or displays. Such work, which requires sustained judgment, has traditionally been tough to automate. By evaluating the extent of autonomy in Claude chat and Cowork to Claude Code, we present that that is beginning to change.

Throughout nearly all sorts of outputs (26 of 31 outputs proven) the extent of AI autonomy is larger on Claude Code than chat or Cowork.¹¹ For instance, conversations producing scripts and code snippets contain 0.53 factors extra autonomy (on common, on the 1-5 scale) when created with Claude Code than conversations producing the identical output on chat or Cowork. Throughout all conversations the typical distinction in autonomy is 0.37 factors, and it has two most important sources.¹²

Roughly two thirds of the distinction is defined by the identical duties being executed with extra delegation on Claude Code. Weblog posts and articles illustrate this: the requests and duties behind them are related on the 2 surfaces, however the best way individuals work with Claude differs sharply. The median chat and Cowork dialog producing a weblog put up or an article includes 13 rounds of back-and-forth, whereas the median blog-producing Claude Code session accommodates a single human immediate. The remaining third displays the completely different mixture of output sorts throughout the 2 surfaces.

One would possibly suspect this distinction merely displays mannequin alternative. Claude Code periods run on probably the most succesful fashions much more typically (54% are served by Opus, in opposition to 10% of chat and Cowork conversations). Nevertheless, the hole persists once we examine conversations served by the identical mannequin. For instance, amongst conversations utilizing Sonnet, Claude Code periods nonetheless present 0.26 factors extra autonomy, suggesting that the product used is probably going extra necessary than the underlying mannequin.

Stepping again from the floor comparability, the output sorts the place customers delegate probably the most are the identical ones that devour probably the most compute: throughout artifacts, imply autonomy and median token use rise collectively (r = 0.68 on chat and Cowork; Appendix Determine A.2).

Claude solutions above the extent it was requested

For every dialog, a classifier estimates two studying ranges—one for the consumer’s immediate, one for Claude’s response—expressed because the years of schooling wanted to know the textual content.¹³ We discover that studying degree varies broadly relying on artifact sort. A mean question leading to an instructional paper would require greater than 16 years of schooling, roughly equal to bachelor’s degree, and 15% of those conversations are at PhD degree or above (20 or extra years of schooling). On the opposite finish of the spectrum are conversations leading to recipes or steerage, the place fewer than 10 years of schooling are required to know the immediate.

Generally, artifact sorts with higher-reading-level outputs even have higher-reading-level prompts (a correlation of 0.87 throughout conversations). Nevertheless, we additionally observe that in nearly each class, Claude’s output is at the next comprehension degree than the immediate, by roughly one 12 months of schooling on common. The hole is widest the place customers describe one thing to be constructed, akin to picture and graphics (+2.6 years), video games (+1.9), and apps and web sites (+1.7). Among the hole might merely be register; prompts are sometimes terse and casual, whereas Claude tends to answer in polished prose. Nevertheless, the hole is close to zero for audience-facing writing (blogs −0.1, tutorial papers +0.0, e-mail +0.3), probably as a result of prompts sometimes draft language or supply materials written in the identical register because the supposed output.

Perceptions

The primary two chapters present how individuals use Claude, however don’t give a lot perception into the methods individuals expertise AI at work—how they anticipate their jobs and workplaces to vary, how they really feel about AI’s present and potential affect, and what they hope for from the know-how. Our interviews with 81,000 Claude users, performed in December 2025 with Anthropic Interviewer, gave a picture: respondents reported giant productiveness positive aspects, but in addition expressed fear about displacement. These worries have been concentrated amongst early-career employees and occupations the place we observe Claude doing probably the most work.¹⁴

In April 2026, we launched the Anthropic Economic Index Survey to construct on this work. The survey permits us to ask individuals straight about their expertise with AI and work, and to discover how responses range with Claude utilization. We hyperlink survey responses to utilization knowledge from mid-Might to early June utilizing privacy-preserving methods. To characterize every respondent’s utilization patterns, we randomly pattern as much as 20 periods per individual inside this time window (throughout Claude.ai, Cowork, and Claude Code, in order that the combo of periods displays every individual’s typical utilization throughout surfaces). We exclude respondents with fewer than 5 periods to scale back sampling noise. Our remaining linked pattern consists of about 9,700 survey respondents.

We discover that the majority respondents anticipate important AI progress over the following 12 months. Whereas individuals’s notion of AI capabilities will depend on their expertise, the place they dwell, and the way uncovered their job is to AI, their expectations in regards to the tempo of future progress are strikingly uniform, in line with a “rising tide,” during which AI capabilities enhance broadly.

Views on what that progress means for their very own careers are much less uniform. Early-career employees report that AI can do the best share of their work and categorical probably the most concern about job loss. But—opposite to a standard concern—the individuals who delegate to Claude probably the most are the most optimistic about their future labor market outcomes, and really feel their abilities are rising in worth. And regardless of (or maybe due to) their proximity to AI’s frontier, the typical respondent’s hopes for the following decade middle not on alternative however on collaboration. They hope AI can protect significant work and automate the drudgery, and that its positive aspects shall be shared broadly.

Who responded to the Financial Index Survey

The Financial Index Survey will not be consultant of the final inhabitants. We attain a random pattern of Claude customers, there could also be choice in who completes the survey, and we filter out rare customers from our evaluation. Determine 3.1 exhibits the occupational mixture of survey respondents (orange) alongside US employment (gray). Laptop and Mathematical occupations are probably the most closely over-represented, making up roughly 30% of survey respondents—akin to their share of Claude utilization, however far above their 4% share of US employment. Administration, at 23% of respondents,¹⁵ can be closely over-represented relative to its 7% employment share, regardless that it accounts for under 4% of periods. This hole is in line with managers utilizing Claude for duties apart from administration itself: within the survey, judgment and administration are named by many respondents (particularly these with extra expertise) as capabilities AI lacks. Bodily occupation classes like Transportation & Materials Transferring, Meals Preparation & Serving Associated, and Development & Extraction are all under-represented within the survey, as they’re in Claude periods as properly.

AI and work duties

Analysis on AI impacts typically focuses on occupational publicity, or what share of duties inside a given job are doable with AI. In prior work, we constructed a measure of noticed publicity, which captures the share of occupational duties we already see being finished with Claude. We in contrast it to a generally used measure of theoretical exposure, or the share of occupational duties that a big language mannequin might theoretically do.

One other solution to perceive occupational publicity is to easily ask individuals how a lot of their job AI is able to doing. We requested respondents what share of their work duties AI might do totally by itself in the present day (hereafter reported publicity), and what share they anticipate it to deal with in 12 months (anticipated publicity), with the choice to pick from 5 bands ranging between “nearly none” and “practically all.” Shut to six in 10 respondents selected the next band for subsequent 12 months than for in the present day. Over a 3rd anticipate AI to have the ability to do most or practically all of their work duties subsequent 12 months (Determine 3.2).

Determine 3.3 compares reported and anticipated publicity to noticed and theoretical publicity. We ask whether or not what individuals report and anticipate AI can do traces up with the noticed and theoretical publicity measures throughout occupations, and whether or not respondents whose occupations rating larger on noticed or theoretical publicity anticipate sooner progress over the following 12 months. On the primary query, the reply is sure: reported publicity (gray dots) is positively correlated with each noticed and theoretical publicity. On the second, the reply is not any: the best-fit traces for reported and anticipated publicity 12 months from now (orange dots) are basically parallel, which means that folks in roles with excessive noticed or theoretical publicity anticipate roughly the identical enhance within the share of their work duties AI can do over the following 12 months as these in roles with much less noticed and theoretical publicity.¹⁷ In different phrases, a software program engineer and a development supervisor anticipate roughly the identical increment of progress inside their occupation.

Additionally it is price noting that reported publicity systematically exceeds noticed publicity. One clarification for that is that not everyone does each job in an occupation, and our survey disproportionately reaches those that use AI extra.¹⁸ Analogously, since theoretical publicity is an higher sure on what is feasible as a substitute of a measure of present use, theoretical publicity systematically overstates reported publicity.

We additionally study how perceptions of AI’s present and future capabilities relate to the traits and utilization patterns of respondents. The left panel of Determine 3.4 exhibits that perceptions of AI’s capabilities are negatively correlated with nation GDP:¹⁹ The typical share of duties individuals report AI can do for them now’s about 10 share factors decrease amongst high-income international locations. This sample is in line with the chance that AI substitutes for a bigger share of the duties that employees in lower-income international locations do day-to-day, even when occupation-level publicity metrics—which are typically larger in superior economies—recommend in any other case. Certainly, the IMF has noted that whereas superior economies face broader AI publicity total, employees in lower-income international locations might have much less entry to the complementary abilities and infrastructure that enable AI to enhance moderately than substitute their work. In earlier work we documented that lower-income economies have a tendency to make use of Claude in additional automated methods even when adjusting for variations in job combine.

The center panel exhibits that reported and anticipated publicity are additionally negatively correlated with years of labor expertise.²⁰ Folks with at the very least 15 years of expertise put that share of duties AI can do roughly 10 share factors decrease than these of their first 12 months of labor. We discover proof that this can be as a result of skilled employees have collected tacit or context-specific experience that’s tough for an AI to imitate. In follow-up questions, we requested individuals what duties they thought AI would by no means be capable to do and why; the most typical responses emphasised that AI lacks the judgment, contextual consciousness, and situational reasoning that their work requires. Respondents, and disproportionately these with at the very least 15 years of expertise, additionally pointed to the relational and interpersonal dimensions of their jobs—constructing belief and managing individuals—as issues AI can’t replicate.

As with occupational publicity to AI, we discover that perceptions about future enhancements in AI capabilities are basically uncorrelated with GDP per capita and years of expertise. The anticipated share of duties that AI will be capable to do in 12 months is uniformly larger than perceptions about AI’s capabilities in the present day.

We subsequent study the connection between how individuals work together with Claude and their present perceptions of Claude’s capabilities. As with previous experiences, we distinguish between “automation” and “augmentation” modes of collaborating with Claude. We determine conversations as automated when Claude is requested to finish a job with little to no enter from the consumer. Concretely, automation share is the share of conversations whose sample is both directive (“translate this doc”) or a suggestions loop (“edit this e-mail…make it extra informal”).²¹

The appropriate panel of Determine 3.4 exhibits that reported and anticipated publicity rise with automation share. This may very well be as a result of delegation is informative about capabilities—individuals who hand over total duties observe straight what AI can full by itself—or as a result of individuals who already consider AI can do their work are probably the most prepared at hand it over. The identical patterns maintain once we substitute automation share with the share of periods dedicated to work duties, or the share performed in Claude Code.²²

AI and jobs

We additionally ask how individuals suppose their jobs will change within the subsequent 12 months. Greater than a 3rd of respondents mentioned it was probably or very probably that obligations would considerably change (for themselves, a peer, a junior colleague, and a senior colleague). 10% rated shedding their very own jobs as probably or very probably. That is barely beneath the annualized hazard fee of shedding a job within the US;²³ nevertheless, since our respondents skew towards data employees in steady employment (a bunch that plausibly faces below-average separation danger at baseline), this may occasionally nonetheless point out elevated perceived danger. When requested an open-ended query about what was driving their forecasts, 38% of the respondents who rated their job loss as probably or very probably attributed their forecasts to AI.²⁴ Notably, respondents have been on common extra anxious about job loss for others than for themselves.²⁵ Respondents have been particularly anxious about job loss for his or her junior colleagues, with over one third stating that the chance of a junior colleague shedding their job within the subsequent 12 months was over 60%. Respondents have been additionally extra involved about job loss (for everybody) in lower-income international locations.

Are individuals who use Claude in additional automated methods additionally extra anxious about shedding work? We study what individuals mentioned about AI’s anticipated affect over the following 12 months on six dimensions of labor: pay, job safety, capability to discover a new job (financial dimensions) and which means, autonomy, and human interplay (intrinsic dimensions); and have a look at how these expectations differ by the automation share of Claude utilization.

Throughout all six dimensions, individuals with the next share of automated periods really feel extra optimistic in regards to the impact of AI on their job outcomes subsequent 12 months in comparison with those that use Claude extra augmentatively. We noticed the biggest results on expectations about optimistic impacts on future pay and skill to discover a job.²⁶

A pure query is why automated utilization and sentiment transfer collectively. It’s doable that this relationship is defined by choice, that the individuals most keen about AI are additionally probably the most prepared at hand over total duties to it. We are able to’t rule this out totally, however these estimates don’t meaningfully change once we management for consumer tenure on Claude.ai—which we will consider as a proxy for enthusiasm, as a result of it captures early versus later adopters.

One other chance is that individuals who use AI in additional automated methods expertise extra of its advantages in the present day. In line with our previous findings, giant majorities of individuals report productiveness positive aspects in velocity, scope, and high quality of their work (86%, 82%, and 69%, respectively), whereas 27% report positive aspects by means of price financial savings on companies they might in any other case need to buy.

Along with important productiveness positive aspects, nearly all of individuals additionally report studying extra with AI (68%) and feeling like AI has made their abilities extra invaluable (57%). Determine 3.7 exhibits how these two outcomes range with the share of automated periods. We see that the share of individuals reporting that AI is rising the market worth of their abilities rises with automation share, whereas the share reporting they study extra is roughly flat.

A generally voiced concern about delegation is that handing total duties to AI means offloading pondering, with positive aspects in output coming at the price of studying and talent atrophy. We don’t see this sample right here: heavier delegators report studying on the identical fee as everybody else. Nevertheless, these are self-assessments, and abilities can erode at the same time as they turn out to be extra invaluable and as somebody experiences studying extra, so the information don’t rule out talent erosion.

How utilization differs between genders

To date we now have explored how utilization patterns relate to expectations and habits. Subsequent, we research who makes use of Claude in varied methods. Probably the most putting variations are by gender. Ladies, who make up solely 12% of our linked respondent pattern, use Claude in another way from males. Even after accounting for occupational variations, they’re marginally much less probably to make use of Claude for work, their share of periods in Claude Code is 0.24 commonplace deviations decrease (6.3 share factors), and their automation share is 0.33 commonplace deviations decrease (7.3 share factors). As a substitute, girls have a tendency to make use of Claude extra iteratively, they usually log extra energetic time on chat than males, a sign of extra collaborative engagement.²⁷

What do individuals hope for from an AI-transformed economic system?

The Anthropic Financial Index Survey surfaces a mixture of optimistic and damaging experiences and sentiments with respect to AI, however we finish the survey on a hopeful be aware. The ultimate open-ended query asks respondents to “dream large: what do you hope an economic system formed by AI seems like in ten years?” We ran every survey response by means of a classifier which tagged responses with related themes. We present the highest 5 mostly cited themes beneath. Extra descriptions of every might be discovered within the Appendix.

The most typical theme expressed was certainly one of AI augmentation of labor. Over half of survey respondents expressed some model of desirous to collaborate with AI on work that feels significant, of wanting their profession to nonetheless matter, and/or hoping that new industries come up and create new job alternatives. Concurrently, simply over half of respondents hoped for AI automation—particularly of the tedious components of their jobs—so they might have extra free time and extra space for which means exterior of labor. The third most typical theme, expressed by about one third of survey respondents, was certainly one of shared prosperity: the hope that the financial positive aspects from AI shall be broadly shared.

Dialogue

AI is diffusing quickly all through the economic system, throughout an rising variety of surfaces, with more and more clever outputs. In earlier AI chat interfaces, utilization was easy, contained within the chat window with out net search, software calls, artifacts, or different affordances. Now, Claude fashions can function autonomously for hours by means of Claude Code and Cowork. As these kinds change, the consumer base is shifting as properly. Early adopters have been extremely technical. Our most up-to-date customers apply Claude to duties that command lower wages within the labor market.

On this report, we took a number of steps towards extra informative measurement. First, we started measuring increasingly more often, processing knowledge in hourly samples. This reveals how the cadences of each day life are etched into our utilization logs and opens avenues for future analysis. Second, we started recording artifacts, or the outputs that folks take away from Claude. These make Claude’s work output extra legible, and present some intuitive patterns.

Lastly, utilization knowledge solely carries a lot data. Our survey allowed us, for the primary time, to ask individuals straight about how they use AI and what they really feel about it. We discovered that our survey respondents use AI for greater than we give it credit score for—they report AI can do the next share of their work than the noticed publicity measure for his or her occupation would recommend. Requested to forecast subsequent 12 months’s capabilities, over 35% predicted that AI would be capable to do most of their work.

Precisely classifying the work that Claude does will stay a transferring goal. For instance, as AI capabilities enhance, AIs might more and more work together and alternate with one another, maybe in methods inscrutable to people or easy classifiers. Finally, Claude’s affect on the economic system shall be seen in financial aggregates like employment and productiveness as a lot as its utilization logs. Nonetheless, AI is more likely to have its earliest impacts within the areas the place it does probably the most work, so shedding gentle on these ever-changing utilization patterns will stay a key solution to inform the general public.

Appendix

Accessible here.

Quotation

@on-line{anthropic2026aeiv6,
        creator = {Maxim Massenkoff and Eva Lyubich and Szymon Sacher and Zoe Hitzig and Shaoyi Zhang and Ryan Heller and Peter McCrory},
        title = {Anthropic Financial Index report: Cadences},
        date = {2026-06-26},
        12 months = {2026},
        url = {https://www.anthropic.com/analysis/economic-index-june-2026-report},
}

Authors

Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller, Peter McCrory.

Acknowledgements

Scott Sales space, Keir Bradwell, Meredith Callan, Dexter Callender III, Boris Cherny, Chris Doenlen, Eleanor Dorfman, Jake Eaton, Evan Frondorf, Deep Ganguli, Romello Goodman, Ankit Gupta, Kunal Handa, Rebecca Hiscott, Andrew Ho, Hanah Ho, Jerry Hong, Saffron Huang, Mo Julapalli, Katie Kennedy, Jennifer Martinez, Miles McCain, Kelsey Nanan, Tyler Neylon, Adnan Pirzada, Dianne Penn, Kerry Persen, Sarah Pollack, Ankur Rathi, Santi Ruiz, David Saunders, Ankit Siva, Michael Stern, Ami Vora, Scott White, Heather Whitney, Kim Withee, Ryan Zauk, Jack Clark.