The Anthropic Economic Index in 2026: What Three Releases Show About AI and Work

Anthropic has not published a new edition of its Economic Index this month. The most recent report, Cadences, is dated June 26, 2026, and the newest folder in the public dataset is the June 26 release. The company’s Economics team page lists a Canada country brief on July 14, followed by other research through the autumn, including two reports on October 16 and one on November 5. None of those autumn reports is an Index edition.

That pause is a good moment to read the three 2026 editions together. Each is a snapshot of how Claude, Anthropic’s AI model, is used, and the snapshots do not all measure the same things.

What does the Index measure?

The Index tracks how Claude is used across the economy. Anthropic samples conversations from its consumer apps and records from its own API, has classifiers sort them so that staff do not read individual transcripts, and maps each one to a task in O*NET, the US Department of Labor’s catalogue of occupations and tasks. The January report describes its Claude.ai sample as mostly consumer use and its API sample as mostly enterprise use. It also defines two recurring terms. Augmentation means a person works with the model, for example by iterating on a draft or asking it to check their work. Automation means the person hands over a task with little back and forth.

What did each 2026 edition add?

The January edition, Economic primitives, drew on one million Claude.ai conversations and one million API records from November 13 to 20, 2025. It introduced five new measures for each conversation, including how much autonomy the AI had and whether the task succeeded.

The March edition, Learning curves, repeated those measures on data from February 5 to 12, 2026. It also asked whether people who have used Claude longer get better results. People who signed up at least six months earlier were about 5 percentage points more likely to have a successful conversation, and about 4 points after controls for task and model. Anthropic notes that survivorship could explain part of this.

The June edition, Cadences, changed the method. It samples every day, and its chapter on outputs covers chat and Cowork conversations from April 10 to June 10, 2026. It added a classifier for what each conversation produced and a survey linked to the usage of about 9,700 respondents.

Measure	January 2026 edition	March 2026 edition	June 2026 edition
Data period	November 13 to 20, 2025	February 5 to 12, 2026	April 10 to June 10, 2026, sampled daily
Share of Claude.ai conversations in the 10 most common tasks	24%	19%	Not in the report text
Share of API records in the 10 most common tasks	32%	33%	Not in the report text
Augmentation and automation on Claude.ai	52% and 45%	Augmentation “increased slightly”; shares in a chart	No overall split in the report text
Automation in API traffic	About three-quarters	”Decreased sharply”; shares in an appendix chart	Not in the report text
Occupations with at least a quarter of tasks seen in Claude use, cumulative	49%	About 49%, “barely changed”	Not in the report text
Share of conversations that are personal use	35% of Claude.ai	42% of Claude.ai	Chat and Cowork: about 35% on weekdays, just under 50% on weekends
Overall task success as judged by Claude	67% on Claude.ai, 49% on API	Not in the report text, which compares user groups only	Not in the report text

Each column is taken from the report linked in its heading. “Not in the report text” means the figure does not appear in the body of the published report, although the dataset may hold the underlying counts.

What changed between the editions?

Consumer use widened. The March report found the ten most common tasks made up 19% of Claude.ai conversations in February, down from 24% in November. Personal use rose from 35% to 42% and coursework fell from 19% to 12%, partly because of school holidays.

Business use did not widen in the same way. The top ten tasks rose from 28% of API records in August 2025 to 32% in November and 33% in February, according to the January and March reports. The March report also describes coding work moving from the chat product to the API.

One figure held still. Counting every task seen across all reports to date, the January edition found that 49% of jobs had at least a quarter of their tasks performed using Claude, and the March report says that estimate “barely changed”. The figure is a share of occupations and says nothing about how many workers are in them. For a planner, that suggests recent growth came mostly from more use of tasks already seen.

Is AI helping people do tasks or doing the tasks for them?

The answer has moved in both directions. In January 2025, 56% of Claude.ai conversations were classified as augmentation and 41% as automation, and by August 2025 automation was ahead. In the November 2025 data augmentation was back in front at 52% to 45%, although the January report still describes the underlying trend as toward more automation. The March report found that long-tenure users were more likely to iterate with Claude and less likely to hand tasks over entirely, which Anthropic says pushes back against its own earlier hypothesis that experienced users automate more.

The June report does not give an overall split. It measures how much decision-making people leave to the model, on a scale of 1 to 5, and finds that Claude Code sessions averaged 0.37 points more autonomy than chat and Cowork conversations.

How does consumer use differ from business use?

In the January data, 74% of API use was work-related against 46% on Claude.ai, and about three-quarters of API interactions were classified as automation against less than half on Claude.ai. Task success, as judged by Claude, was lower on the API at 49% against 67%, which the report suggests may reflect the benefit of being able to correct the model in a conversation.

This is the part of the Index most relevant to job design. API use is programmatic, meaning software sends the task to the model, usually with less back and forth. That traffic also includes developers testing and building, so it signals where automation is being tried without proving that a company has deployed it. A related Anthropic paper, Labor market impacts of AI, published March 5, 2026, combines usage data with estimates of what language models could do in theory into a measure it calls observed exposure. The same paper found no systematic increase in unemployment for highly exposed workers since late 2022.

What do the October and November reports add?

Three related reports came out after the June edition. The AI expectations gap, published October 16, compares a Morning Consult survey of 10,980 US adults with an internal poll of 1,334 Anthropic staff. US workers expected AI to be able to do 28% of their current work by 2030 on average, and Anthropic staff expected 75%.

How employment is changing at Anthropic, also October 16, reports that the junior share of the company’s technical hires fell from 27% to 9%, and that by August hours worked by AI agents outnumbered employee hours six to one. The first figure is a share of hiring. The second compares agent running time with employee hours estimated at 40 a week, so it does not measure equal amounts of work. Both come from one company that builds the tools, which limits how far they apply elsewhere.

AI for scientific discovery, published November 5, analyzed 207,049 research sessions from August 2 to September 30, 2026. Its scenario for current capabilities, if widely adopted, has AI taking on 27% to 52% of the researcher time needed for a new study, with a median of 38% across fields. The report presents that as a scenario.

What are the limits?

The Index covers one vendor. Its API figures cover traffic sent directly to Anthropic and, as the June report states, exclude Claude delivered through Amazon Bedrock or Google Cloud Vertex. A company that relies on another vendor’s assistant will not see its own pattern of use in these numbers.

The measures also rely on the model’s own judgment. Classifiers built on Claude assign the task and the interaction type, and success is Claude’s assessment of the conversation. The January report calls its classifiers “directionally accurate” and warns that people choose which tasks to bring, so success rates may be higher than they would be across all work.

The survey in the June report has its own limit. Anthropic says it is “not representative of the general population”: about 30% of respondents were in computer and mathematical occupations, which make up 4% of US employment. With that caveat, 10% of respondents rated losing their own job in the next year as likely or very likely.

How can a workforce planner use it?

Use the Index as evidence of where AI is being applied to work at the level of the task. Because it is mapped to O*NET, its tasks can be lined up against the tasks in your own job descriptions. Tasks that the March report finds in API traffic, such as customer service and sales outreach, are ones developers are already running through software, so they are reasonable first places to look when you review a role. If people with six months or more of use succeed more often on the same tasks, a rollout plan that budgets for licenses without time to learn will likely understate what is needed.

The Index cannot tell you how many roles to change. It does not see your tools or your task mix, and it has no view of whether work done with AI met your quality bar. Those answers come from your own data, and the Index is best used as an outside reference for testing them. When the next edition arrives, check first whether it restores the measures the June report left out.

JobRoute uses the June 2026 release as one input to its task exposure scores. Where the Index observes AI in use on an O*NET task, the score reflects how much use that task gets compared with other observed tasks, and where it has no observation a documented heuristic sets the score, as the methodology page explains.