How it works

What the numbers mean

The main number for a job is the share of its working time that is within reach: time spent on desk tasks short enough for the best AI models to finish on their own, at a chosen reliability, at a chosen date. It is a statement about what models could do if the work were set up for them, with the instructions, files and tools in front of them. It is not a forecast of jobs lost, of savings, or of how fast any company will change. Those depend on decisions no data set can see.

Next to it, every job shows observed use: Anthropic's measure of how much of the job people already hand to Claude at work. The two answer different questions. Within reach counts whole instances of a task a model could finish alone. Observed use counts real use, including help on part of a task, which is why it can run higher.

Below the line of reach, every job shows a measured line of use: at each release of the Anthropic Economic Index since February 2025, the share of Claude use that went to the job's tasks. Each release samples Claude.ai conversations and ties each to an O*NET task; this site sums the shares of a job's tasks, splits a task several jobs share evenly among them, and leaves out the conversations tied to no task. The older releases name tasks in earlier O*NET wording, so they are matched to this site's statements by text, then through O*NET's task ids, then through O*NET's 2010-to-2019 occupation crosswalk by shared words (every such match is listed with its score in the project's files): 96% of each older release's task-tied use matched, and all of the newest. Anthropic's method changed between releases (the sample, the classifier, and the June 2026 file's rounding of each task to a hundredth of a percent), so a job's line starts afresh at each change of method rather than joining across it. It measures where Claude's use goes, not how much of a job's time it covers, and it is Claude.ai only. The figures are in the open data.

Every job page also shows Anthropic's rating of its physical work, published September 30, 2026: Claude Opus 5, searching the web, rated each physical O*NET task by the kind of setting in which a robot it could cite does most of the task (a cell built for the machine, a facility kept to a standard, or any setting), and weighted the tasks by its own estimate of their share of the time into a robot exposure index from 0 to 3. The page shows the rating as Anthropic's, with the robot behind each rated task and its source. It does not change this site's estimates: the share within reach still counts no body work, because a robot shown somewhere is not a robot in a given workplace, and cost, setup and safety decide that.

Sources

• Tasks. O*NET® 31.0 Database, U.S. Department of Labor, Employment and Training Administration: 923 occupations with task data and 18,838 task statements, with how often each is done, how important it is and the share of workers who do it. CC BY 4.0.
• Task lengths. A Stratus estimate. Three models from three companies (Gemini 3.8 Flash, GPT-5.6 Luna and Claude Haiku 4.5) each estimated a simple, a typical and a demanding instance of all 17,578 distinct task statements under one written definition. The typical length is the median of the three. Checked against GDPval, OpenAI's set of 220 real professional tasks whose length the professionals recorded: the estimates ran long by a factor of 1.2 at the median, so every length is divided by 1.2.
• Kinds of work. A Stratus estimate, on two footings while the task-by-task review is under way. Reviewed jobs: each task's time split among five kinds of doing. The rest: whether each task could be done entirely through a computer, judged by a model when the lengths were estimated, with people work picked out by its leading verb. An unjudged task counts as body work, never as desk work.
• Model ability. METR, Task-Completion Time Horizons of Frontier AI Models (Time Horizon 1.1), read September 19, 2026 and checked again November 10, 2026. metr.org
• Frontier estimate. Epoch AI, Epoch Capabilities Index (ECI), the scores file (fetched November 11, 2026), fitted to METR's horizons across the 20 models both measured: an estimate for the newest models, not a measurement. epoch.ai. CC BY 4.0.
• Observed use. Anthropic Economic Index, labor market impacts files (published March 5, 2026, from Claude usage in August and November 2025). CC BY 4.0.
• Use over time. Anthropic Economic Index, all six releases (February 10, 2025 to June 26, 2026; the June 2026 file holds two monthly windows, so seven points), Claude.ai task shares. CC BY 4.0.
• Robots. Anthropic, “What work can robots do?” (Legate-Yang and Massenkoff, September 30, 2026): robot exposure ratings for 18,796 O*NET 29.3 tasks and 923 occupations, matched to this site's tasks by statement (7,522 of the 7,594 physical tasks). CC BY 4.0.
• Pay and headcount. U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, national estimates, May 2025. Cost of an hour from BLS Employer Costs for Employee Compensation, June 2026.
• Industries. The same BLS survey's national industry-specific estimates, May 2025: which occupations each sector, subsector and industry group employs, how many people and at what pay.
• States and metro areas. The same survey's state, metropolitan and nonmetropolitan area estimates, May 2025, across all industries.

How a job's time is split across its tasks

O*NET asks workers how often they do each task, in seven answers from "yearly or less" to "hourly or more", and what share of them do it at all. Each answer is turned into times a year (1, 5, 25, 100, 250, 750 and 2,000), averaged over the answers, multiplied by the share of workers who do the task and by its typical length, and capped at one working year. Each task's share of time is its part of the total. This is an estimate: O*NET does not record hours, and the answers are ranges.

When a task comes within reach

METR measures the time horizon of the best models: the length of task, timed by how long a skilled person takes, that a model finishes half the time (even odds) or four times in five (four in five). Its latest measurement is Claude Mythos Preview (early), April 2026: 2.2 work days at even odds and 3.1 h four in five. Nothing newer is in METR's published results as of November 10, 2026. METR's time-horizons page has said since September 8, 2026 that it is no longer actively updated, so a newer measurement may appear in one of its reports instead; each report is read as it appears, and none since April 2026 states one. METR's June 2026 report on OpenAI's GPT-5.6 Sol gave about 11 hours at even odds by its standard method and said none of its figures is a robust measurement, so it is not used here. The report.

Today's estimate carries METR's own recent trend (frontier models since 2023, doubling every 129 days) forward to the date checked: 2.7 work weeks at even odds and 2.4 work days four in five. The four-in-five length is the even-odds length divided by 5.62, the ratio in METR's latest measurement. Every figure on this site labeled now is this projection, projected from METR's April 2026 measurement and Epoch's November 2026 index, and says so where it is shown.

The frontier estimate: METR's page stopped updating in September 2026, so the newest models' horizons are read from Epoch AI's Capabilities Index, which scores every frontier model within days of its release on one scale built from fifty-odd benchmarks. Across the 20 models both METR and Epoch measured, 4.7 index points go with a doubling of METR's horizon (R² 0.98). On that line Claude Fable 5.2, November 2026, at 174.2 points, would finish tasks of about 2.7 work weeks at even odds, with an 80% range of 4.8 work days to 1.9 work months. The line is extended past the range both measured, and METR says its own suite cannot measure above sixteen hours, so this is an estimate and not a measurement. Today's estimate above carries that point forward on METR's recent trend; METR's trend alone would give 3.9 work days at even odds at the date checked. Only a measurement METR publishes scores the record, and the fit is redone whenever Epoch's file changes (the decision and the candidates considered are in the project's frontier note, read November 11, 2026).

From today the length grows at one of three paces: the long-run pace (doubling every 188 days, METR's whole record since 2019, the default), the recent pace (every 129 days) and a slowdown (once a year).

Instances of a task vary. A task's typical length sits between a simple and a demanding instance, read as the 10th and 90th percentiles of a log-normal spread, taken from the median model. A task comes within reach a piece at a time: its share within reach is the part of its time spent on instances no longer than the horizon. Long instances carry more of the time, so this counts time, not instances.

Three kinds of work

Every page adds work up into three kinds, and only desk work can come within reach. For a job that has been reviewed task by task, the three are the five kinds of doing in the next section, added up. For a job not yet reviewed, each task is one kind as a whole, by this rule.

Body work needs hands, a place or a presence in the room. People work is desk work whose main action is dealing with people: a task counts as people work when one of its leading verbs is on this list:

advise, arbitrate, coach, collaborate, comfort, confer, console, consult, counsel, delegate, demonstrate, direct, discuss, encourage, entertain, escort, facilitate, greet, hire, host, instruct, interview, lead, liaise, lobby, mediate, meet, mentor, moderate, motivate, negotiate, persuade, preach, present, recruit, represent, sell, solicit, supervise, teach, testify, train, tutor, visit, welcome.

Neither counts as within reach here, whatever its length. The rule is deliberately simple so anyone can check it; it will count some tasks as people work that a model could help with, and miss some people work that opens with another verb. Everything else is desk work.

The task-by-task review

A review of every task, one at a time, began on September 30, 2026. It reads each task statement with its occupation and answers, against a written rubric, what the doing consists of as a person does it today. It does not judge what AI can do: model ability stays in METR's numbers, so the review holds across model releases. The largest occupations go first.

So far: 402 of 923 occupations and 8,645 of 18,838 tasks, holding 93% of the jobs BLS counts. As of November 10, 2026.

Each task's time is split, in tenths, among five kinds of doing:

Only the documents and systems parts of a task can come within reach, by the task's length. A task that is 30 percent documents can be at most 30 percent within reach. Each reviewed job page also splits its share within reach between the two, because they ask different things: documents can be handed to a model, while work inside a business system needs the model connected to that system. Conversation that is short enough by length is shown beside them and never counted.

The review judges four more things about each task. Whether the law reserves it to a licensed, certified or sworn person (a licensed act). What a mistake costs: small, costly or severe. How far it follows a script: a set procedure, a procedure with judgment, or open-ended. And whether it deals with outside people or protected records. None of these changes the share within reach. They say what handing a task over would ask.

The reading is done by Claude Fable 5.1, an AI model from Anthropic, one occupation at a time, against the same rubric and the same worked examples, and a script refuses any occupation that is incomplete or inconsistent. Each session re-judges a sample of earlier tasks without seeing the earlier answers: across 300 tasks so far, the leading kind matched 94% of the time and the whole split 64%. The judgments are estimates: another careful reader would split some tasks differently.

What it changes: across the occupations reviewed so far, 30% of working time is within reach now, against 34% by the whole-task rule they stood on before (weighted by the people in each). Industries, areas and company scans use the reviewed figures as they arrive, so their numbers move a little with each batch until the review is complete.

What changes for the people

For each job reviewed task by task, the site splits the working time three ways at the end of 2028, on the long-run pace at four in five. What stays with people is the conversation, in-person and hands-on part of every task. What becomes checking is the documents and systems part within reach by then, in tasks where a mistake costs something real or is severe, or where a licensed person must sign: someone still has to check or sign the result. What could be handed over with lighter checks is the documents and systems part within reach in tasks that are small if wrong and need no license. The rest of the desk work is not yet within reach by then. What to learn is one line for each pattern the judgments show, largest first: documents work of a tenth of the time or more (briefing a model and checking its drafts), systems work of a tenth or more (running AI inside the business systems), any checking (reviewing work you did not do yourself), licensed tasks over 5% of the time, severe stakes over 5%, conversation and in-person work of 15% or more, hands-on work of 15% or more, outside people over 15%, protected records over 10%; and, where a graded attempt fell short on a criterion, that criterion. No model is asked; the rules run on the task-by-task judgments.

Industry and area pages, company pages, the full report and What changes for workers add these maps up, weighted by how many people do each job, and count how many people have each skill among the first two things to learn for their job. It says which parts of the work change, never whether anyone keeps a job.

Hours and cost

A company scan multiplies each role's share within reach by its paid hours (2,080 per person a year) and headcount. Value uses the company's pay where given, otherwise the BLS median for the matched occupation, times the cost of an hour: BLS reports that wages were 70.0 percent of what private employers spent per hour worked in June 2026, so the default multiple is 1.43. The value within reach is the cost of that work, not a saving.

Titles are matched to occupations in the browser, against O*NET's lists of titles people actually use. When the same words are one occupation's own title and another's alternate, the occupation whose title it is wins (Sales Engineer is Sales Engineers' own title). Otherwise equal matches go to the larger occupation (where BLS counts several occupations only together, as it does the three kinds of buyer, the larger as of May 2016, the last count that told them apart). A match with a close rival in a different occupation, or one made through only part of the title, is marked Check, with the reason beside it.

Industries

BLS counts the people in each occupation in each industry, for 20 sectors, 85 subsectors and 247 industry groups (NAICS). An industry's share within reach is its occupations' shares weighted by those counts; its hours and value follow the company scan, using the pay the industry pays each occupation. The same counts, scaled to a headcount, make the typical company each industry page can scan.

BLS publishes a few occupations only combined, such as buyers and purchasing agents (13-1020) for O*NET's three kinds of buyer. One rule places every such case: within a group of related codes, the O*NET occupations BLS does not publish belong to the one combined code BLS publishes there, and where that leaves a choice, nothing is guessed. Twelve combined codes are placed this way. The people BLS counts only in "all other" groups, which have no task lists, are left out, and each industry page says how much of its employment the task data covers: 98% for the median industry.

Seen in Claude use splits today's share within reach by whether Anthropic saw people bring each task to Claude at work in 2025. It counts one AI product, so it shows where use has begun, not where it has not.

Each industry and area page also shows what handing the work over would take, for the part of its people in occupations reviewed task by task, with that part stated: what their work is made of across the five kinds of doing, and the share of their working time within reach split into documents and data, business systems, and conversation, which is shown and never counted. The rest of the people join the split as the review reaches their occupations.

States and metro areas

The same survey counts the people in each occupation in every state and territory, 393 metropolitan areas and 137 nonmetropolitan areas, across all industries. An area's numbers are built exactly as an industry's, from those counts and the pay each area pays. BLS leaves out small counts it cannot publish, so task data covers a little less of an area's people than of an industry's; each area page says how much.

Each area page also runs the area's whole workforce through the company scan: every occupation BLS publishes there that the site has task data for, at the local headcount and pay, as one roster. Its totals are the area page's; the scan adds every occupation, the tasks behind the hours and the downloads. Occupations BLS does not publish for an area are not in its roster.

Industry and area pages also say what changes for their people: each reviewed job's own reskilling map (what stays with people, what becomes checking, what can be handed over, and the first two things to learn), weighted by how many people do that job there, at the end of 2028 on the long-run pace at four in five.

Companies

Each company page is read from the company's latest annual report on EDGAR: the headcount it states, the share in the U.S. when it says, and its workforce as a blend of the industry staffing patterns above, with a one-sentence reason per share that can be checked against the filing. The estimates on the page are the industry estimates weighted by those shares. The filing's sentences that name AI are counted and quoted, never judged, and the same count is kept for every annual report since fiscal 2021 of every company in the Index, read or not, so each company page shows its line by year and what annual reports say about AI adds them up by year and sector. A sentence counts when it names AI as written: “AI” in capitals, “artificial intelligence”, “generative AI”, “gen AI”, “machine learning” or “large language model”. The reading is done by Claude Fable 5.1 (1,367 companies) and Claude Fable 5.2 (60 companies) under one written rubric, each company page naming the model that read it, and a sample of earlier readings is judged again blind each session by the model then reading, once the file passes fifty companies, with the agreement rate published here: the latest, on November 11, 2026 by Claude Fable 5.2, re-read 5 and matched the headcount 100% of the time, the leading pattern 80%, and every share within ten points 80%.

Headcount by filing year comes from each year's annual report, linked on the page. A licensed data service helps locate the filings and the figures; each session a sample of those figures is looked for in the text of the filing it cites, and the share found is published here: 100% of 50 points checked on November 11, 2026. The Index covers every US-listed company with 1,000 or more employees in its latest annual report (1,835 companies on October 1, 2026); 1427 have been read so far, and the ranking is partial until all are read. Each sector has its own page with its ranking and how much of it is read, and every company page shows its rank among the companies read in its sector. When a 10-K incorporates its workforce section by reference from the annual report to shareholders, that exhibit is fetched from the filing's index on EDGAR and read with the 10-K, and the page says so. When a company files its next annual report, that report is read in turn and the earlier reading is kept as it was; the company page then lists each annual report read, priced with the same estimates so the rows differ only by what the reports say. The figures come from the filing and from BLS, never from a roster, and no page says what a company will do. Every company read is also a row of open data, with its identifiers, stated headcount, estimates and filing, on the data page (CC BY 4.0).

Worked examples: the Work Bench

The Work Bench takes common tasks from the task-by-task review, in order of the desk hours they carry across the US among tasks that are mostly documents work: the 500 that carry the most first, shorter ones first among them. Once all 500 have an attempt, the share of their desk hours whose latest attempt passed is published as the Work Bench number, with its date. For each, a realistic brief and its material are written for the example and are invented, never a real company's; a rubric is written and timestamped before any attempt; the model then makes one attempt from the brief and the material only, and that first output is published unedited; the grade comes last, criterion by criterion against the rubric, as passes, passes with named fixes, or fails, with notes. Each example's brief, material and rubric were written by Claude Fable 5.1, which also made the first attempt and graded it, in that order in one session, so each page shows the three timestamps and says so. At each later model release the same brief, material and rubric are run again: Claude Fable 5.2 made its attempts in a later session, from the brief and the material only, and graded them against the unchanged rubrics, and every earlier attempt stays on its page. A person's check of a sample of grades will be published beside them. So far 20 examples: 9 passed, 11 passed with named fixes, 0 failed. An example is one attempt at one task on one date. It is not a statement that a job, or the task in a given workplace, can be done by AI, and it does not change any number on the site: the shares within reach still come from task lengths and METR's measurements. When a new model is available the same briefs are run again, and earlier attempts stay on each page.

How the numbers move

Every job and industry page shows how far its share within reach has come since GPT-4: at each frontier measurement METR has published since March 2023, the share within reach at four in five using the best model measured by that date and today's task data, then today's estimate. It is a backcast. It shows how fast the models have moved, not what anyone forecast at the time.

From this release on, every data build also keeps a dated record of every job's and industry's numbers, including the projections for the end of 2027, 2028 and 2030. As new measurements arrive, those records will show how each estimate changed, and in time how the projections compared with what the models actually did.

Limits

• METR's tasks are mostly software, research and reasoning work. In July 2025 METR found time horizons 40 to 100 times shorter for work done by operating a computer screen, such as clicking through web pages. Desk tasks that depend on operating software a model cannot reach may come within reach later than shown. Reviewed job pages show how much of their share within reach is work inside business systems.
• Within reach assumes the work is set up for the model: context, files, systems access and a way to check the result. Most work is not set up that way yet.
• Task lengths, time shares, the kinds of work (reviewed task by task, or by rule for jobs not yet reviewed) and today's frontier are estimates. Ranges are shown where they matter.
• An occupation is an average. Any one company's version of a job can differ a lot from it.
• An industry is an average too: a typical company is the industry's staffing, not any real company's.
• Nothing here is advice about any person's job or any company's staffing decisions.

Terms

Within reach

Work a model could finish on its own at the chosen reliability, if the work is set up for it: the instructions, files and tools it needs are in front of it. Within reach is not the same as done. Whether a company hands the work over, and how fast, is a separate question.

Four in five

The length of task the best models finish four times out of five. It is the default here, because work a company hands over has to come back right most of the time. Today's estimate: 2.4 work days.

Even odds

The length of task the best models finish half the time. It runs about 5.6 times the four-in-five length. Today's estimate: 2.7 work weeks.

Pace

How fast the length of task models can finish keeps growing from today's estimate. Three are offered: the long-run pace (doubling every 188 days, the default), the recent pace (every 129 days) and a slowdown (once a year).

Reliability

How often a model has to finish a task for the task to count as within reach. Four in five is the default; even odds is the rate METR headlines.

Time horizon

METR's measure of AI ability: the length of task, timed by how long it takes a skilled person, that a model finishes at a given rate. It has doubled every four to seven months since 2019.

Frontier estimate

Since METR's time-horizons page stopped updating (September 2026), the newest models' horizons are read from Epoch AI's Capabilities Index, which scores every frontier model on one scale from fifty-odd benchmarks. Across the 20 models both METR and Epoch measured, 4.7 index points go with a doubling of METR's horizon; Claude Fable 5.2 scores 174.2, or about 2.7 work weeks at even odds on that line, range 4.8 work days to 1.9 work months. The line is extended past the range both measured, so this is an estimate, not a measurement, and only a METR measurement scores the record.

Today's estimate

METR's latest measurement is Claude Mythos Preview (early), April 2026: 2.2 work days at even odds and 3.1 h four times in five, and METR has published nothing newer. Epoch's Capabilities Index scores the newer models on one scale; a fit of METR's horizons on that index across the 20 models both measured puts Claude Fable 5.2 (November 2026) at about 2.7 work weeks at even odds, with a range of 4.8 work days to 1.9 work months. Today's figure carries that point forward on METR's recent trend to the date checked: 2.7 work weeks at even odds and 2.4 work days four times in five. An estimate, not a measurement.

Long-run pace

The length models can finish doubles every 188 days, the pace of METR's whole record since 2019. The default here.

Recent pace

Doubles every 129 days, the pace METR measures for frontier models since 2023.

Slowdown

Doubles once a year, about a third of the recent pace. It shows the line if the gains from reasoning and tools run out.

Range

The low end assumes the slowdown pace and the high end the recent pace. The middle figure assumes the long-run pace.

Desk work

Work a person could do entirely through a computer. The only kind this site counts as within reach. In jobs reviewed task by task it is the documents and systems parts of each task; in the rest, whole tasks judged desk work.

People work

Dealing with people: conferring, negotiating, supervising, teaching, selling and the like. Counted here as staying with people. In jobs reviewed task by task it is the conversation and in-person parts of each task; in the rest, whole tasks whose leading verb is on the list on the method page.

Body work

Work that needs hands, a place or a presence in the room. Length says nothing about it, so it never counts as within reach. In jobs reviewed task by task it is the hands-on part of each task.

Reviewed task by task

Every task of this job has been read one at a time against a written rubric: how its time splits among five kinds of doing, whether a licensed person has to answer for it, what a mistake costs and how far it follows a script. The review describes the work as people do it today, not what AI can do. It is working through all 923 jobs, largest first.

Made of

How a task's time splits among five kinds of doing: documents and data, business systems, conversation, in person, and hands-on. Judged task by task, in tenths. Only the first two can count as within reach.

Documents and data

Reading, writing, calculating, analyzing, planning, designing, coding, reviewing. The test: someone working from another city could do it with only the files and a way to send the result back. Counts as desk work.

Business systems

Entering, looking up, updating, scheduling, ordering, posting, or watching and controlling through software. The result is a changed record or an operated system, so whoever does it needs access to that system. Counts as desk work.

Conversation

Live talk with people by phone, video or chat, and meetings that could be held that way. Counted as people work: this site leaves it with people, however short the task.

In person

Work that needs this person there: a class, a bedside, a sales floor, a crew on site. Also the person's own learning, which nobody can do for them. Counted as people work.

Hands-on

Handling things, tools, machines, vehicles or bodies, or being at a place to see, check, guard or move. Counted as body work.

How scripted a task is

Set procedure: the same steps each time, fixed by rules or a checklist. Procedure with judgment: a known process, with calls to make when cases vary. Open-ended: no script; the person has to diagnose, design, create, weigh or persuade. Judged task by task.

What a mistake costs

Small if wrong: rework, a delay, a corrected record. Costly if wrong: money, a customer, a contract, a compliance finding. Severe if wrong: someone could be hurt, or the loss is large and hard to undo. Judged task by task.

Licensed act

The law requires the person doing or approving this task to hold a license, a certification or a sworn office for it. Whoever prepares the work, a licensed person stays answerable for the result.

Protected records

The task works with records about a person that the law protects: health, financial accounts and credit, tax, student, personnel and pay, criminal or immigration records, or privileged legal material.

Outside contact

The task deals directly with people outside the organization: customers, clients, patients, students, the public, suppliers or officials.

Observed use

Anthropic's measure of how much of a job's work people already hand to Claude at work, from Claude usage in August and November 2025, published March 2026. It counts help with part of a task, which is why it can be higher than the share within reach.

Share of Claude use

Of the Claude.ai conversations Anthropic sampled and tied to a work task, the share that went to this job's tasks, at each release of the Anthropic Economic Index. A task several jobs share is split evenly among them. It measures where Claude's use goes, not how much of the job's own time it covers.

Robot exposure index

Anthropic's measure, from 0 to 3, of how far robots have been shown doing a job's physical tasks: each task scores 1 if a robot does most of it in a setting built for the machine, 2 in a facility kept to a standard, 3 in any setting, 0 otherwise, weighted by the task's share of the working time. Rated by Claude Opus 5 with web search, September 2026.

Task length

How long a person takes to do one instance of the task. Estimated by three AI models from three companies and checked against GDPval, where professionals recorded how long real work took. Shown as a typical instance, with a simple and a demanding one either side.

Share of time

How much of the working year a task takes. Estimated from O*NET's survey answers (how often workers do the task and how many of them do it) times the task's length.

Paid hours

2,080 a year for a full-time person, the basis BLS uses to turn hourly pay into annual pay.

Hours within reach

Paid hours times the share of each role's time within reach, times headcount, added up over the team.

Cost of an hour

Pay plus benefits and payroll taxes. BLS reports that wages were 70.0 percent of what private employers spent per hour worked in June 2026, so pay is multiplied by 1.43. You can change it.

Value within reach

What the hours within reach cost the company at the cost of an hour. It is the cost of the work, not a saving: what a company does with the time is its own decision.

Occupation

One of the 923 jobs O*NET data describes task by task, from the US Department of Labor's standard list.

O*NET data

The US Department of Labor's database of occupations, the tasks that make them up, and how often and how widely each task is done. This site uses version 31.0.

Match

The occupation a job title was matched to, using O*NET's lists of about 60,000 titles people actually use. High means an exact or near-exact title. Check anything marked Check or Pick one: the reason is beside it.

Core task

A task O*NET marks as central to the job rather than supplemental.

Median pay

Half of the people in the occupation earn more and half less. From the BLS Occupational Employment and Wage Statistics survey, May 2025.

NAICS

The North American Industry Classification System: the codes the US government uses to sort businesses into industries. Two digits name a sector (42 is wholesale trade), three a subsector, four an industry group. BLS publishes a few groups combined, with codes like 4230A1.

Seen in Claude use

The part of today's share within reach that sits in tasks Anthropic saw people bring to Claude at work in 2025 (Anthropic Economic Index). The rest is within reach in tasks not yet seen there. Other AI tools are not counted, so this understates use; it shows where use has started, not where it has not.

MCP

Model Context Protocol: the open standard ChatGPT, Claude and other assistants use to call tools outside themselves. An MCP address is where the assistant sends those calls.

Connector

Claude's name for an outside service it can call through MCP. ChatGPT calls the same thing an app.

Changes

Every projection is kept in a dated snapshot and scored against METR's later measurements on the record; every change is also on updates, with a feed.

• November 10, 2026, that evening. The Index now has 1,427 companies read from their own annual reports, 60 more, from Zumiez, Warby Parker, Etsy, Affirm, Hims & Hers and GATX to Devon Energy, Match Group, CoreWeave, Bloom Energy, Oscar Health, Klaviyo and Otter Tail. These are the first read by Claude Fable 5.2, and each company page names the model that read it. A seventeenth blind re-read, the first in which one model re-read another's work, agreed on the headcount for all five companies and on the leading pattern and every share within ten points for four; both differences were judgment calls settled with the earlier readings standing (Principal Financial, Trimble). The filing's own count replaced the first figure collected thirteen times, among them Zumiez's 9,000 with its part-time staff against 2,400 full-time and Warby Parker's 4,036 against 2,275. 408 companies wait.
• November 10, 2026, later that day. The Work Bench has its first repeat. The twenty worked examples were attempted again by Claude Fable 5.2, the model this site's reading is now done by, from the same briefs and material and against the rubrics written on October 1 before any attempt. Each new attempt was written before its rubric was opened and is published unedited above the earlier one. Graded by itself, strictly to each rubric's stated evidence, the newer model passed 9 and passed 11 with named fixes, with none failed; Claude Fable 5.1 had graded its own attempts 19 and 1. Most of the named fixes are for length: the newer attempts are 45% longer in total and run past the word limits several rubrics set, and one disputes a rubric's arithmetic and is graded as not meeting it. Each model grades its own work, so the difference comes from the grader as well as from the work; every example page shows both attempts and both sets of notes, and a person's check of a sample is still to come. Company readings and worked examples made from this point name Claude Fable 5.2 (their pages carry the date in UTC, November 11); everything read before keeps the name of the model that read it, on its own page and in the method.
• November 10, 2026. Today's estimate now starts from a newer point. Epoch AI's Capabilities Index scored Claude Fable 5.2, released November 9, 2026, at 174.2 (its own interval 170.1 to 178.8), above GPT-6 Astra's 169.2. On the line fitted across the twenty models both Epoch and METR measured (4.7 index points to a doubling of METR's horizon, R² 0.98), that reads as about 108 hours of work at even odds, with an 80% range of 38 to 325 hours. It is an estimate, not a measurement: the line is now extended about sixteen index points past the range both measured, the score is two days old and rests on the benchmarks Epoch has run so far, and METR's own trend alone would give 31 hours. Today's estimate moved from about 73 hours to about 109 at even odds, and from 13 to 19 hours four times in five. The share of paid hours within reach now across the US moved from 27.8% to 30.0%, the end of 2027 from 34.2% to 34.9%, the end of 2028 from 36.0% to 36.2% and the end of 2030 from 36.5% to 36.6%; accountants and auditors moved from 33% to 45% now, software developers from 38% to 48%, registered nurses from 35% to 37%, and customer service representatives stayed at 64%. Every figure labeled now says it is projected from METR's April 2026 measurement and Epoch's November 2026 index; the method page has the fit and its limits, and the record keeps every earlier snapshot.
• November 8, 2026, that evening. The Index now has 1,367 companies read from their own annual reports, 119 more, from AMC Entertainment, International Flavors & Fragrances, Prologis and Robinhood to Carlyle, Rocket Lab, Reddit, Tripadvisor, GitLab, Evercore, Enphase, Trinity Industries, Granite Construction, Alnylam and Jazz Pharmaceuticals, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A sixteenth blind re-read of five earlier readings agreed on the headcount for all five and on the leading pattern and every share within ten points for three; the two differences (GoDaddy's technology staff at web hosting rather than software, and Teladoc's lead at care delivery rather than software) were settled as worked examples with the earlier readings standing. Twenty headcounts the Index had taken from the wrong sentence were corrected to what the filings state, among them AMC Entertainment at 33,311 where the Index showed its 2,931 full-time associates, International Flavors & Fragrances at 21,500 where it showed its 3,000 research employees, Trinity Industries at 6,110 where it showed its 2,650 in the United States, Granite Construction at 5,800 where it showed 2,500 salaried, Citi Trends at 4,700 where it showed 2,500 full-time, Seneca Foods at 2,995 where it showed its 4,015 seasonal workers at the harvest peak, and Carlyle at 2,500 where it showed the 700,000 people its portfolio companies employ. Carlyle Secured Lending was set aside unread as a fund with no employees of its own. 468 companies wait.
• November 8, 2026, later that day. The Index now has 1,248 companies read from their own annual reports, 60 more, from Interactive Brokers, Roblox, EOG Resources, BioMarin, Lazard, Vornado, Primoris and Genworth to Lucky Strike Entertainment, 1-800-Flowers and Star Group, with a fifteenth blind re-read that agreed on the headcount and the leading pattern for all five and corrected UGI's reading to show its midstream business apart from its gas utility. 587 companies wait.
• November 8, 2026, later that day. Every figure labeled now, on job, industry, area and company pages, in the scan and the reports, the workbook and the slides, the assistant tools and the API, now says what it is projected from. METR's time-horizons page stopped updating in September 2026, so today's estimate starts from Epoch AI's Capabilities Index, fitted to METR's horizons across the twenty models both measured (4.7 index points to a doubling of the horizon, R² 0.98): GPT-6 Astra, September 2026, reads as about 51 hours at even odds on that line, range 22 to 124 hours, carried on METR's recent trend to the date checked, against 31 hours on METR's trend alone. An estimate, not a measurement: the share of paid hours within reach now across the US moved from 22% to 28%, the end of 2028 from 35% to 36%, the end of 2030 not at all; the method page has the fit and its caveats, and every earlier snapshot stands on the record.
• November 8, 2026. The Index now has 1,188 companies read from their own annual reports, 99 more, from CME Group, Simon Property Group, Apollo Global Management and Lyft to Instacart, Roku, GameStop, Planet Fitness, The Trade Desk, Crown Castle, Kodak, Kelly Services and PBF Energy, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A fourteenth blind re-read of five earlier readings agreed on the headcount and the leading pattern for all five and on every share within ten points for two; the one correction was ADT's reading, which now shows its 4,000 customer care professionals as a counted support operation beside its security sales, installation and service staff, and the other two differences (Acushnet's footwear share and Jones Lang LaSalle's on-site facilities staff) were settled as worked examples with the earlier readings standing. Twenty headcounts the Index had taken from the wrong sentence were corrected to what the filings state, among them Kelly Services at 4,900 internal staff where the Index showed the 375,000 workers it placed, TrueBlue at 3,500 where it showed the 123,000 people it connected with work, Associated Banc-Corp at 4,000 where it showed a dollar figure, Glacier Bancorp at 4,188 where it showed a share count, GameStop at 15,000 where it showed the 16,000 of its seasonal peak, Boot Barn at 12,700 where it showed its 8,900 part-time Partners, and Strategic Education at 6,134 where it showed its 3,683 full-time employees. Plains GP Holdings was set aside unread as a second listing of Plains All American's workforce. METR checked on its own site the same day: its results file still ends at Claude Mythos Preview (April 2026), none of its reports since states a new time-horizon estimate, and its time-horizons page has said since September 8, 2026 that it is no longer actively updated, which the method and the record now say.
• October 19, 2026, still later that night. The Index now has 1,089 companies read from their own annual reports, 116 more, from Blackstone, KKR, Pinterest and Snap to Palantir, Coinbase, Moderna, Norwegian Cruise Line, Toll Brothers, Hasbro and Six Flags, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A thirteenth blind re-read of four earlier readings agreed on the headcount for all four and on the leading pattern and every share within ten points for three; the one difference corrected Vertiv's reading, which now shows its thermal management products as heating and cooling machinery beside its power equipment. Fourteen headcounts the Index had taken from the wrong sentence were corrected to what the filings state, among them Norwegian Cruise Line at 44,500 where the Index showed its 5,300 shoreside employees, Caleres at 10,000 where it showed 5,000 full-time, Covenant Logistics at 4,700 where it showed its 2,900 drivers, Knife River at 5,298 where it showed its 6,900 peak-season count, Six Flags at 4,225 full-time where it showed 91,000 seasonal and part-time employees, and Matthews International at 5,500 where it showed the 10,000 of a venture it part-owns. Four more listed companies with no workforce of their own were set aside unread, their Index counts being their managers' or their parent's.
• October 19, 2026, later that night. The Index now has 973 companies read from their own annual reports, 120 more, from Vistra, Okta, MongoDB and Sempra to PulteGroup, NVR, Public Storage, GoDaddy, Teradyne, Harley-Davidson, Vail Resorts and Addus HomeCare, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A twelfth blind re-read of five earlier readings agreed on the headcount and the leading pattern for all five and on every share within ten points for four; the one difference corrected McKesson's reading, which now reads the oncology and multispecialty practices, the medical-surgical distribution and the prescription technology business apart from the drug distribution instead of folding them into it. Twenty-five headcounts the Index had taken from the wrong sentence were corrected to what the filings state, among them Addus HomeCare at 49,948 where the Index showed its 5,290 full-time employees (44,658 of its caregivers are part-time), Sempra at 15,938 where it showed Oncor's 5,600, Murphy USA at 16,900 where it showed 5,900 full-time, Helmerich & Payne at 15,700 where it showed its 6,200 in the United States, Valmont at 10,791 where it showed its 6,294 in the United States and Columbia Sportswear at 9,620 where it showed its 5,730 retail employees.
• October 19, 2026, late night. The Index now has 853 companies read from their own annual reports, 139 more tonight, from Lincoln National, Clorox, Snowflake and Illumina to Airbnb, United Rentals, Datadog, Biogen, Zoom, Shopify, Canadian Pacific Kansas City and American Water, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. An eleventh blind re-read of five earlier readings agreed on the headcount and the leading pattern for all five and on every share within ten points for three; the two differences (Apple's split among its product business, its stores and its services, and Fiserv's reading wholly at transaction processing) were settled as worked examples and the earlier readings stand. Twenty-nine headcounts the Index had taken from the wrong sentence were corrected to what the filings state, among them United Rentals at 28,500 where the Index showed its 8,200 salaried employees, Canadian Pacific Kansas City at 19,479 where it showed 7,000 Canadian union members, and Five Below at 24,600 crew where it showed 7,800 full-time.
• October 19, 2026, night. The Index now has 714 companies read from their own annual reports, 119 more tonight, from Phillips 66, Lennar, eBay and Entergy to Kinder Morgan, CrowdStrike, Occidental, Nasdaq, Weyerhaeuser and ConocoPhillips, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A tenth blind re-read of five earlier readings agreed on the headcount and the leading pattern for all five and on every share within ten points for four; the one difference (DuPont, split across chemicals, machinery and plastics from the product lines its filing describes) was settled as a worked example and the earlier reading stands. Three utilities' stated service companies now sit at management of companies as the earlier readings do (Entergy, FirstEnergy, Eversource), and twenty-two headcounts the collector had taken from the wrong sentence were corrected to what the filings state, among them Parsons at over 21,000 and Solventum at 20,000.
• October 19, 2026, evening. The Index now has 595 companies read from their own annual reports, 119 more today, from Gilead, Netflix, Moody's and Regeneron to Con Edison, D.R. Horton, Expedia, Aflac and Take-Two, each with its headcount as the filing states it, its U.S. share where stated, and its workforce blended from the site's industry patterns with a reason a reader can check against the filing. A ninth blind re-read of five earlier readings agreed on the headcount and the leading pattern for all five and on every share within ten points for three; one difference was settled as a worked example (a title insurer's offices and underwriter split near evenly when the filing counts neither), and the other corrected an earlier reading whose sales force sat at a broader industry than the site carries, after which seven more readings were moved to the industry group that covers their work (Cigna, Rocket, Lam Research, Dover, Savers, McKesson, Cencora) and Crown Holdings was re-split between its can plants and its strapping business.
• October 19, 2026, afternoon. The scan's title index now carries all 2,000 judged job titles: a fourth blind audit of 30 earlier judgments agreed on the occupation, on the fair reading and on the check for all 30, then the last 1,200 titles were judged in six batches, from architectural associate and captain through surgeon and wholesale representative, each with a reason a reader can disagree with and the alternatives a careful coder would allow. On the half held out for testing, the matcher now names the judged occupation or a listed alternative 82.1% of the time (73.8% exactly), against the 95% the program's launch gate asks; the misses are the class rules the matcher still lacks, and they are the next work on it. The judging is complete, so the Index is the program the deposits go to next.
• October 19, 2026. The task file now covers 402 of 923 occupations (8,645 tasks, 93.5% of US employment): a tenth blind audit of 30 earlier judgments agreed on the exact mix 53% of the time and on the leading kind 100%; no judgment was wrong by the rubric, but the audit showed one recurring shape, treatment on a patient's body, that the rubric had settled only by convention, so the rule is now written down (treating by hand is hands-on work for the whole task), eight earlier judgments were revised to it and three worked examples were added; then 12 more occupations were read, among them interpreters and translators, nurse anesthetists, sales engineers, photographers, surveyors, transportation security screeners and concierges. The Index now has 476 companies read from their own annual reports, 50 more today, from Raymond James and Norfolk Southern to Intuit, Live Nation, American Electric Power and Newmont; an eighth blind re-read of five earlier readings agreed on the headcount for all five and on the leading pattern and every share within ten points for four, and the one difference corrected an earlier reading (Stanley Black & Decker, whose power tools are machinery, not metalworking machines). The scan's title index gained 200 more judged job titles (800 of 2,000) and now carries every judged title: on the half held out for testing, the matcher names the judged occupation or a listed alternative 85.2% of the time.
• October 18, 2026, night. The task file now covers 390 of 923 occupations (8,381 tasks, 93.1% of US employment): a ninth blind audit of 30 earlier judgments agreed on the exact mix 67% of the time and on the leading kind 93%, no judgment was changed and three worked examples were added, then 14 more occupations were read, among them tax examiners and revenue agents, procurement clerks, tree trimmers, actors and orderlies. The Index now has 426 companies read from their own annual reports, 50 more tonight, from Arrow Electronics and Grainger to MercadoLibre, Exelon and Workday; a seventh blind re-read of five earlier readings agreed on the headcount and the leading staffing pattern for all five and on every share within ten points for three, with the two differences settled as worked examples. The scan's title index gained 200 more judged job titles (600 of 2,000): on the half held out for testing, the matcher now names the judged occupation or a listed alternative 83.9% of the time.
• October 18, 2026, evening. Every company the Index has read now has a report to keep, $29: the page as a document with what it does not show, the share within reach at every date on three paces and two reliabilities, a typical workforce of the company's shape at its stated headcount role by role and task by task (the industries' staffing, never its roster), where to start, what changes for the people, every sentence in its filing that names AI, its headcount by year, its peers, the method and a citation line, to print or save as a PDF; Walmart's report is the free sample. Every job page now ends with the personal report's offer, opening on that job. The task file covers 376 of 923 occupations (8,123 tasks, 92.6% of US employment): an eighth blind audit of 30 earlier judgments agreed on the exact mix 53% of the time and on the leading kind 93%, four judgments were corrected and four worked examples added, then 13 more occupations were read, among them glaziers, telemarketers, clergy, hospitalists, urologists and tool and die makers. The Workforce AI Index has read 376 of its 1,835 companies: a sixth blind re-read of five earlier readings agreed on headcount, leading pattern and every share, and 49 more annual reports were read, from Sunbelt Rentals, Howmet Aerospace, Fastenal, BlackRock, PayPal, Casey's, CSX, Garmin, Motorola Solutions, Xerox and XPO among others. The job titles' third blind audit agreed on the occupation for 28 of 30 titles.
• October 3, 2026, early morning. The task file now covers 363 of 923 occupations (7,866 tasks, 92.1% of US employment): a seventh blind audit of 30 earlier judgments agreed on the exact mix 57% of the time and on the leading kind 87%, one judgment was corrected and four anchors added. The Index holds 1,835 companies after a repair of its listing rule: nine companies that had stood behind a listed note or preferred share (Southern Company, Fifth Third, Assurant, Kinder Morgan, Unum, DTE Energy, Liberty Latin America, CMS Energy and Kemper) now stand under their common stock, and 30 more companies are read from their annual reports, 327 in all, with a fifth blind re-read agreeing on every headcount. The job titles' second blind audit agreed on the occupation for all 30 titles.
• October 2, 2026, late evening. The task file now covers 351 of 923 occupations (7,602 tasks, 91.6% of US employment): a sixth blind audit of 30 earlier judgments agreed on the exact mix 63% of the time and on the leading kind 93%, one judgment was corrected to the rubric (an invoice goes to the customer) and two worked examples added, then 12 more occupations were read, among them marriage and family therapists, fraud examiners, school psychologists, labor relations specialists, credit analysts and financial risk specialists. The Workforce AI Index has read 297 of its 1,827 companies: a fourth blind re-read of five earlier readings agreed on headcount, leading pattern and every share, and 49 more annual reports were read, from U-Haul, Visa and Kraft Heinz to Uber, Broadcom, Dick's Sporting Goods, Union Pacific and General Mills; in six of them the stated total differed from the sentence the collector had picked, and the reading follows the filing. The Titles program has judged 400 of its 2,000 titles, with its first blind audit (the same occupation 93% of the time, the same or a listed alternative 100%) and the matcher's second score on the harder held-out half (the judged occupation or a listed alternative 82% of the time, against the 95% gate on the method page).
• October 2, 2026, late. The task file now covers 339 of 923 occupations (7,351 tasks, 91.1% of US employment): a fifth blind audit of 30 earlier judgments agreed on the exact mix 63% of the time and on the leading kind 90%, six judgments were corrected to the rubric's settled cases and four worked examples added, then 13 more occupations were read, from web developers and database administrators to structural iron and steel workers and general internal medicine physicians. Every number was rebuilt and a new snapshot dated. The Workforce AI Index has 248 companies read from their own annual reports, 80 more today from Exxon Mobil and Chipotle to NVIDIA and Alaska Air Group, after a third blind re-read of five earlier readings agreed on the headcount, the leading pattern and every share.
• October 2, 2026, night. What annual reports say about AI: the sentences that name AI in every annual report since fiscal 2021 of the companies in the Workforce AI Index, counted by one rule from the filings on EDGAR, by year, by sector and by company, with a lookup of any company in the Index. Each company page shows its own line by year beside the Index's median report, each sector page its sector's, and the quarterly report's draft the Index's. The counts are open data, in the quarterly release and in the API. Naming AI is not using it: a sentence counts whether it describes a product, a risk, a rule or a plan, and nothing is judged. The personal report now links each task's worked example where there is one, as the scan does.
• October 2, 2026, evening. The company scan now sets the roles scanned beside the public companies the Workforce AI Index has read: the share of paid hours within reach now and by the end of 2028, on the company pages' basis, among all the companies read, those whose largest staffing pattern is in the same sector, or the same pattern, with the companies nearest it listed. A scan started from an industry page opens on that industry's peers, and the full report carries the comparison as a workbook sheet and a slide. Each company's figures come from its own annual report and BLS staffing for its industries, never from its roster. Wherever a task in a scan has a worked example on the Work Bench, in Where the hours come from, Where to start and Pro's plan, it links to the graded attempt, and the workbook carries the link. The method now says that nothing newer than April 2026 is in METR's published results, and names METR's June 2026 report on GPT-5.6 Sol, whose figure METR does not consider a robust measurement.
• October 2, 2026. A new home page shows a job's 40-hour week as forty squares, hour by hour, at each date to 2030: what AI could take on, what becomes a person checking its work, the desk work AI can't do yet and what stays with people, then the same for a typical company, a map of what a company can do with freed hours (buy fewer hours, or put the same people on more and better work) and what stands behind every number. Every job page opens with a one-sentence summary and the same week picture, and any two jobs can be compared side by side. The scan opens with its answer in a paragraph, the header searches every job, company, industry and place, company pages show their figures as cards and bars, the pricing page sets the tiers side by side with pictures of the report, and every company page and worked example has its own share picture. Pages that offer an area's or an industry's report say what its roster holds (the occupations BLS publishes with task data, most of its people), and the home page says the value of the hours within reach is their cost in pay, benefits and payroll costs, not a saving.
• October 2, 2026. Every state, metro and nonmetro area's workforce as BLS publishes it (the occupations with task data, most of its people) 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.

Questions about the method: info@stratussc.com. Back to the start