Updates

The Workforce Scan grows every few days: more occupations reviewed task by task, more companies read, more worked examples, and every new measurement of what AI models can do. This page says where each part stands and what changed when. Follow it with the feed.

Where things stand

AI ability

Claude Mythos Preview (early), measured by METR in April 2026, is the latest measurement; checked again October 1, 2026.

The task-by-task review

326 of 923 occupations and 7,098 of 18,838 tasks, 91% of the jobs BLS counts, checked by 120 blind re-judgments so far.

The Workforce AI Index

168 of the 1,827 US-listed companies with 1,000 or more employees read from their annual reports. The first quarterly report is issued at the close of 2026.

The Work Bench

20 worked examples, each a rubric written before one graded attempt.

Measured Claude use

7 points from the Anthropic Economic Index, the latest for May 2026, on every job page.

Robots

Anthropic's robot exposure rating of September 30, 2026, on every job, industry and area page.

Open data

Jobs, industries, areas, the backcast, Claude use and the Index as CSV under CC BY 4.0; the first frozen quarterly release at the close of 2026.

Next

• More occupations reviewed task by task and more companies read, every few days.
• When METR publishes a measurement newer than April 2026: every estimate refreshed, and the earlier projections scored against it.
• The week of December 28, 2026: the Index's first quarterly report and the first frozen open data release.

Every change

• October 2, 2026. Every state, metro and nonmetro area's whole workforce now runs through the company scan from its area page, at the headcount and pay BLS reports there, and the full report on it is the regional report, with a sample for Worcester, MA, and a page for workforce boards and colleges. Industry and area pages say what changes for their people (how much of the working time stays with people, becomes checking a model's work or could be handed over by the end of 2028, and the skills most of them would learn first), added up for the US and every state on What changes for workers; company pages carry the same three shares, the full report adds up the skills across a roster, reviewed job pages answer what to learn, and the open data files carry the figures. Every job, industry, area and company page offers a small card to embed elsewhere with the credit built in. A page for lenders and insurers. The Work Bench shows its progress toward the 500 tasks that carry the most documents work, which it now takes first, and is ready to re-run its examples at the next model release and to publish a person's check of a sample of grades. The API gained industries and areas, and what changes for the people on jobs, industries, areas and companies. The same whole-workforce scan works for an industry nationwide, from the last choice of its typical-company menu. Every job page can look up how many people do the job in any state or metro area and at what pay, and company pages list the read companies with the same leading staffing pattern.
• October 1, 2026, night. The company scan says why each match needs a look (another occupation fits the title as well, or only part of the title matched), and the full report's workbook carries the same column. When the same words are one occupation's own title and another's alternate, the occupation whose title it is now wins: Nursing Assistant, Sales Engineer, Facilities Manager, Food Service Manager and Order Clerk match those occupations. The personal report has a sample to read before buying, shows other names each occupation goes by when you choose yours, shows every figure on a phone, adds the share of each task within reach by the end of 2028, and says on paper whose job it reads and when. The record sets each worked example's grade beside what the estimates projected for its task on the day of the attempt. The API's company answers include the headcount by filing year, each with its filing, and the readings of earlier annual reports. Pro's scan results show, role by role, what stays with people, what becomes checking a model's work, what can be handed over and what to learn first, and every full report carries the same in its workbook and a slide in its deck. METR checked: no measurement newer than April 2026.
• October 1, 2026, evening. Every job page shows a measured line of Claude use beside its estimated line of reach: the share of Claude.ai use that went to the job's tasks at each of the Anthropic Economic Index's releases since February 2025, with its rank among the jobs and how many of its tasks showed up, starting a new line where Anthropic's method changed. The figures are a new open data file. Pro subscribers get a plan for each saved roster: its reviewed tasks in four waves by a stated test (documents first, then work inside business systems, then work a person must sign, then what comes within reach by the end of 2028), with a status and a target date for each piece kept in their browser and carried into the slides. Every job, industry and area page shows Anthropic's new robot exposure rating of its physical work (published September 30), with the robot behind each rated task on job pages, as Anthropic's rating beside this site's estimates, which still count no body work; the index is a new column in the jobs table and the open data files, and every job is ranked by it. Every job is also ranked by its share of Claude use. Every job reviewed task by task shows what changes for the people in it: what stays with people, what becomes checking the machine's work, and what to learn. Pro plans keep an outcome log: the hours each piece of work gave back, as the company enters them, beside the projection, in the workbook too, and each saved roster downloads as a file to keep or move to another browser. A portfolio view sets every saved roster side by side, ranked by the hours that can start now. Anyone can read their own job against its occupation's tasks, with a pasted job description if they like; the full personal report is on sale at $19. Pro by the year is now $2,990, two months free (it was $1,188, the price of four months, which made the monthly price look inflated); nobody had bought it. The API is open to Pro subscribers under its license, and the Prove-it Lab explains how a company can test today's models on its own tasks. This page of updates began, with a feed. The record opened: what each dated snapshot projected for the length of task AI models finish, on each pace, to be scored against the first METR measurement made after it. Actors, dancers, musicians and singers, and disc jockeys outside radio now take their own pay, the hourly median BLS publishes for a full year of 2,080 hours, instead of a broader group's annual pay; the value of hours within reach moved with it in the industries and areas that employ them. METR checked: no measurement newer than April 2026.
• October 1, 2026, afternoon. The Workforce AI Index as open data: every company read, with its CIK, ticker and exchange, the headcount its annual report states, the share of working time within reach now and to 2030, the leading staffing pattern, the count of its filing's sentences that name AI and the filing itself, as a CSV on the data page under CC BY 4.0. When a company files its next annual report, that report is read in turn and the earlier reading is kept, and the company page lists each one. The Work Bench opened: twenty worked examples of common tasks, each a rubric written before one unedited attempt by Claude Fable 5.1 and its grade, linked from the tasks they show on each job page. Every industry and area page now shows what handing the work over would take for the part of its people in occupations reviewed task by task (about nine in ten for most). Job, industry, area and company pages answer the questions people ask about them, from their own numbers. Pro subscribers can save rosters and open them again when the data changes, to see what moved role by role. METR checked: no measurement newer than April 2026.
• October 1, 2026, later. The Workforce AI Index reached 168 companies read, the pilot of fifty and the largest of the rest by headcount, with a page for every sector and a sector rank on each company page. IBM and Wells Fargo were read again with the annual reports their 10-Ks incorporate by reference, so IBM now carries a stated headcount. The second blind re-read of five earlier readings matched the headcount, the leading pattern and every share within ten points 100% of the time, and a second sample of fifty headcount history points was found in the filings 100% of the time. A royalty trust whose last headcount dated from 1997 left the Index universe under a class rule, which now holds 1,827 companies. The task-by-task review reached 326 occupations and 7,098 tasks, 91% of the jobs BLS counts, after its fourth blind audit (the same leading kind 90% of the time, the whole split 63%; two judgments corrected to the settled cases). Across the US, 21% of paid hours are within reach now and 35% by the end of 2028, unchanged. METR checked: no measurement newer than April 2026.
• October 1, 2026. The Workforce AI Index opened: seventy companies read from their own annual reports, the pilot of fifty and the first twenty of the Index by headcount, each with the reading, the estimates that follow, the filing's own sentences about AI, and its headcount by filing year with every filing linked; the Index lists the 1,830 US-listed companies with 1,000 or more employees it will cover, and says the ranking is partial until they are read. The task-by-task review reached 313 occupations and 6,828 tasks, 90% of the jobs BLS counts, after its third blind audit (the same leading kind 97% of the time, the whole split 77%; four judgments corrected to the worked examples). Across the US, 21% of paid hours are within reach now and 35% by the end of 2028, unchanged. METR checked: no measurement newer than April 2026.
• September 30, 2026, night. The task-by-task review reached 289 occupations and 6,318 tasks, 89% of the jobs BLS counts, after its second blind audit: 30 earlier tasks judged again without the file in view matched the leading kind 100% of the time and the whole split 67%, and two judgments were corrected to the worked examples. Across the US, 21% of paid hours are within reach now and 35% by the end of 2028, against 21% and 36% before this batch. Equal title matches among occupations BLS counts only together (the three kinds of buyer) now go to the larger as of May 2016, the last count that told them apart. METR checked: no measurement newer than April 2026.
• September 30, 2026, evening. The task-by-task review reached 167 occupations and 3,723 tasks, 78% of the jobs BLS counts, and its first blind audit: 30 earlier tasks judged again without the file in view matched the leading kind 93% of the time and the whole split 73%; four judgments were corrected to the rubric. Across the US, 21% of paid hours are within reach now and 36% by the end of 2028, against 22% and 37% before this batch. Pro can now be paid by the month, and a page for private equity says how to use the scan in diligence. METR checked: no measurement newer than April 2026.
• September 30, 2026. The task-by-task review began: the first occupations now split each task's time among documents, systems, conversation, in-person and hands-on work, with what handing each task over would ask. Their shares within reach moved with it (see the review). Company scans show the same split for the roles reviewed so far, and the full report lists where to start. Actors, dancers, musicians and disc jockeys now show BLS's own count of people rather than none. METR checked: no measurement newer than April 2026.
• September 29, 2026, later. Every state and metro area. How far each job, industry and area has come since GPT-4, and a dated record of every estimate from here on. Industries added: every sector, subsector and industry group, with a typical company to scan. Twelve occupations BLS publishes only combined now take the combined code's pay instead of a broader group's; their shares within reach are unchanged.
• September 29, 2026. First release. METR checked: no measurement newer than April 2026.