A report is what its authors saw in the data. A file listing is what they were willing to hand over, and the two are rarely the same size. The Anthropic Economic Index has both: six reports on anthropic.com and a public repository, which the dataset card on Hugging Face describes as data on how AI is used in real-world tasks. On 8 October 2026 we read that card and every release note under it, listed the repository through the Hugging Face API, and pulled rows over HTTP. The listing came to 86 files and 763.4 MB.
This piece reports what is in those files. It does not say what they mean for anyone’s job.
What files are in the Anthropic Economic Index dataset?
The repository has eight folders. Six are dated releases, from release_2025_02_10 to release_2026_06_26, and the card lists six numbered reports beside them. The other two are labor_market_impacts and robot_exposure. Of the 86 files, 43 are CSV.
The releases have grown quickly. The first holds 14 files and 5.2 MB. The latest holds three files and 296.5 MB, which is 38.8 percent of every byte in the repository. The main Claude.ai file went from 18.9 MB in the September 2025 release to 94.1 MB in January 2026, 103.3 MB in March and 219.2 MB in June.
The robot_exposure folder is a different kind of dataset. Its README says Claude Opus 5, with web search, rated physical O*NET tasks for exposure to robots, and it lists 923 occupations and 18,796 tasks. Only 105 tasks sit in the top tier, the one for unstructured settings such as a private home. That is a model’s rating of what robots could do, which is a separate thing from our count of how many humanoid robots shipped in the 2026 filings.
One practical finding came first. The Hugging Face viewer does not work on this repository. On the day we checked, the datasets-server is-valid endpoint returned false for preview, viewer, search, filter and statistics, and the splits endpoint returned an error. The error names a file path taken from the card’s own config block. That block points the raw_1p_api split at the release_2025_09_15 folder, the file is in release_2026_01_15, and the path as written returned HTTP 404 when we requested it. So we read the CSV files directly, with range requests for the large ones. When we asked where the humanoid robot dataset from the World Humanoid Robot Games was, the question was whether files existed. Here they exist and the preview in front of them is switched off.
What is one row of the Economic Index data?
No row is a conversation. The documentation for the 26 June 2026 release defines a row as one metric value for one geography and one category node, and gives ten columns: date_start, date_end, geo_id, geo_level, category_name, hierarchy_level, metric_id, value, node_name and node_external_id. There is no user, no prompt and no reply anywhere in the schema.
The category is one of four: overall, onet for O*NET work activities, request for a topic hierarchy of what people ask, and soc_occupation for the Standard Occupational Classification. The geography is GLOBAL, a country code or a subregion code.
There are two files in the release. The claude_ai file covers Claude chat and Cowork on the Free, Pro and Max plans, broken out by country and subregion. (What those plans cost a team is a separate matter, which we went through in our piece on AI subscription costs and per-seat pricing.) The 1p_api file is global only, and the documentation describes it as “Anthropic 1P API calls, excluding Claude Code.”
The date range has a seam in it. The three releases before June each carry one week in the file name: 4 to 11 August 2025, 13 to 20 November 2025 and 5 to 12 February 2026. The June release covers April and May 2026 as calendar months. Anthropic’s June 2026 report says the pipeline now samples a slice of conversations every day, where earlier reports drew on seven-day samples. Anthropic calls this privacy-preserving telemetry. We have looked at telemetry from the receiving end before, in our Local Agent Telemetry Report 2026 on what sixteen coding agents send.
We fetched the first 2 MB of the 219.2 MB Claude.ai file. That slice held 16,680 complete rows: 7,104 global, 5,530 country and 4,046 subregion. Every date_start was 2026-04-01 or 2026-05-01. This is a slice of under 1 percent of the file. The rows are in no order we could see, so we do not treat it as a fair sample of the whole.
What did we compute from the Economic Index files ourselves?
We ran three checks. The first uses complete files. The release_2025_03_27 folder has task_pct_v1.csv and task_pct_v2.csv, which its README calls version 1 and version 2 of the task percentage data. Each has two columns, task_name and pct. We loaded both whole, summed pct, sorted by pct and compared the task names.
Version 1 has 3,514 tasks and version 2 has 3,365. In both, pct sums to 100.0. The same task is top of both lists: modifying existing software to correct errors, adapt it to new hardware or improve performance. Its share is 4.79 percent in version 1 and 6.65 percent in version 2. The ten largest tasks add up to 21.15 percent in version 1 and 25.10 percent in version 2. Only 2,781 task names appear in both files, with 733 in version 1 alone and 584 in version 2 alone, so the two lists are not the same set measured twice.
The second check found a trap. The same folder has automation_vs_augmentation_v1.csv and v2.csv, six rows each, columns interaction_type and pct. The directive row reads 22.56 in version 1 and 29.42 in version 2. But the six version 2 rows sum to 100.00 and the six version 1 rows sum to 84.21. The README does not say where the other 15.79 points went. We did not compute a change between two columns that stand on different bases.
The third check is on the latest release, and it is rows from a slice, not a ranking. From the first and last 2 MB of the 1p_api file, the global overall rows for April 2026 give use_case_work_pct 82.19, use_case_personal_pct 14.7 and use_case_coursework_pct 3.11. Those three sum to 100.00. In the Claude.ai file, the global overall row we caught for use_case_work_pct is May 2026, at 43.36. The months differ, so we draw no gap from the pair. One like-for-like pair did turn up: for May 2026, the Life, Physical, and Social Science occupation group has a pct of 4.51 in the Claude.ai file and 4.29 in the API file, 0.22 points apart.
What can the Anthropic Economic Index not tell you?
Start with volume. The January and March 2026 documentation says the raw file contains raw counts and percentages. The June documentation lists percentages, means and one index, and no count metric. None of the 54 metric names in our slices contained the word count. From the June files alone, a share cannot be turned back into a number of conversations.
Absence is ambiguous too. Cells are published only above aggregation thresholds and a sample floor per geography, and the documentation says a missing row means the cell was not published, “not necessarily that the value is zero”. Published values are rounded to two decimal places.
The labels are a model’s. The June report says of its classifiers that “transcripts are only read by another instance of Claude”. Several metrics are described in the documentation as estimates, among them the time a task would take a person without AI and the years of education it needs.
Coverage has edges. Every row measures use of Anthropic’s own products. The API file has no geography and leaves out Claude Code. The licence is stated two ways: the card’s metadata tag says mit, and its text says the data is CC-BY and the code MIT.
The June documentation says future release schedules are still to be announced. The repository was last modified on 30 September 2026 and its newest dated folder is still release_2026_06_26. If the next folder carries June, July and August in the same ten columns, the monthly series grows from two calendar months to five, and the one-week windows before it become the only part a reader has to stitch.
