Jobs Demand Index

Leading indicators

AI Frontier

Job postings, official forecasts and pay are lagging indicators: they move after employers have decided. The leading indicators are how fast AI can do professional work and how fast organizations are putting it to use. This page tracks both beside the index's hiring data, so you can see disruption while there is still time to act.

Leading and lagging, on one chart

Each line is divided by its own level in May 2024 and drawn on a log scale, so a line that doubles climbs the same height wherever it starts. The units differ, so read the slopes, not the levels.

Leading: capability
17 hours

of skilled-human work: the length of task the best AI agent measured by METR completes on its own half the time (Claude Mythos Preview (early), released Apr 7, 2026). In Apr 2025 it was 2 hours (o3).

Leading: use
3.2 quadrillion

tokens a month processed by Google alone in May 2026, about seven times a year earlier, by Google's count.

Lagging: hiring
−9.4%

US job postings, all jobs, on May 2024 (Indeed Hiring Lab index, data to Nov 13, 2026). Software development postings: +18.1%.

Change since May 2024, log scaleCapability and use (leading) against job postings (lagging)
Show the numbers
MeasureMay 2024LatestMultipleSource
Task length the best AI agent completes half the time7 min17.4 h (Apr 7, 2026)149×METR, by model release date
Tokens a month processed by Google9.7 trillion3.2 quadrillion (May 2026)330×Google, I/O 2026
Canadian businesses using AI (past 12 months)6.1%19.2% (Q2 2026)3.1×Statistics Canada
US job postings, all jobs (index)115.9105.0 (Nov 13, 2026)0.91×Indeed Hiring Lab
US job postings, software development (index)69.181.6 (Nov 13, 2026)1.18×Indeed Hiring Lab

Sources: METR Time Horizon 1.1; Google, Google I/O 2026 keynote; Statistics Canada, AI use by businesses, Q2 2026; Indeed Hiring Lab job postings index (CC BY 4.0). Multiples are JDI's arithmetic on the published figures.

The leading lines have multiplied many times over in two and a half years. Hiring has moved in single or low double digits, up in some fields and down in others. That distance is the warning. Employers adopt the tools first, then redesign the work, and only then change how many people they hire and for what. By the time a change shows up in postings or in an official forecast, the decisions behind it were made months earlier.

Frontier log

Verified milestones, newest first. Claims are reported as claims, with the dispute where there is one. Each entry says what it means for work in one plain sentence.

  1. Expert timelines

    Dwarkesh Patel shortens his timeline for automated AI research

    In an essay titled "The avalanche that hit math is coming for AI development", Patel writes that AI as good as the best human AI researchers is a couple of years away, possibly months. His median is 2028. Earlier in 2026 he put superintelligence at least most of a decade away.

    What it means for work. A prominent long-timelines voice moved sharply after the math results. Plans that assume a decade of slow change are the ones to stress-test.

    Source: Dwarkesh Patel, Dwarkesh Podcast, Nov 13, 2026.

  2. Capability

    New top score on GDPx: 16.4%

    Claude Fable 5.2 was listed first on SemiAnalysis's GDPx leaderboard at 16.4%, a week after launch, ahead of Claude Opus 5.5 at 12.0%.

    What it means for work. The share of multi-hour professional tasks an AI finishes alone went from 12.0% to 16.4% in the benchmark's first week.

    Source: ArtificialWatch, Nov 10, 2026, reading the GDPx leaderboard.

  3. Capability

    SemiAnalysis launches GDPx: the best model scores about 12%

    GDPx tests AI agents working alone for up to eight hours on tasks in 11 fields, including chip design, hardware, finance, legal and software engineering. SemiAnalysis reported that the best model scores about 12%, and that under strict pass-or-fail grading every model scored zero on finance and legal.

    What it means for work. AI already completes about one multi-hour professional task in eight with no human help. The score covers defined tasks; judgment, accountability and trust, the blind spot of AI, sit outside it.

    Source: SemiAnalysis, Nov 2, 2026.

  4. Usage

    Google's model APIs pass 32 billion tokens a minute

    Alphabet said its first-party model APIs process nearly 32 billion tokens a minute, more than four times a year earlier.

    What it means for work. Businesses are building AI into their own systems at scale. API traffic is work being routed to models, not people trying a chatbot.

    Source: Alphabet, Q3 2026 earnings call, Oct 28, 2026.

  5. Math and science

    OpenAI claims an existence proof for four-dimensional gauge theories, including QCD

    OpenAI said its models produced existence proofs for interacting quantum field theories with gauge fields and fermions in four dimensions, a family that includes QCD, the theory of the strong force. A committee led by the physicist David Gross is examining the work before the proof is released. As of mid-November no independent check or peer review had been published, and the mass-gap part of the related Millennium Prize problem remains open. This is a claim, not a verified result.

    What it means for work. If it holds, AI is producing frontier mathematics, not only routine analysis. Until independent experts confirm it, treat it as a claim.

    Source: Interesting Engineering, Oct 26, 2026; Tech Insider, Nov 2026.

  6. Math and science

    OpenAI publishes about 720 AI-written math manuscripts

    OpenAI released a public repository of manuscripts from an unreleased internal model working on open problems. On Nov 17, 2026 the repository listed 721 manuscripts in 373 families; about 42% of top-line results have machine-checked Lean proofs. Mathematicians including Frank Calegari and Tristan Buckmaster have called many papers poorly written and some wrong, and OpenAI did not release the prompts.

    What it means for work. Research output that took careers can now be produced in bulk. Checking it, judging which results matter and taking responsibility for them stays with people.

    Source: OpenAI math repository on GitHub (Apache 2.0); Let's Data Science and Tech Insider coverage, Oct to Nov 2026.

  7. Math and science

    OpenAI claims a proof of blow-up for the 3D Navier–Stokes equations

    OpenAI said an internal system of about 10,000 agents, running for 88 hours, produced a 166-page proof with a Lean formalization that a smooth fluid at rest can develop a singularity under a smooth external force. The Clay Mathematics Institute said on Sep 11 that the problem has apparently been settled, with its review deliberately unhurried. The unforced case remains open, and mathematician Tristan Buckmaster disputes how credit was handled.

    What it means for work. One of the seven Millennium Prize problems drew a claimed answer from 88 hours of machine time. Mathematicians say the proof teaches little so far: understanding and explaining it remains human work.

    Source: Clay Mathematics Institute, Sep 11, 2026; NPR via GPB, Sep 22, 2026; The Tufts Daily, Sep 24, 2026.

  8. Math and science

    Terence Tao warns good open problems are being used up

    Tao wrote that fruitful open problems are being depleted in a non-renewable fashion as AI tools solve them faster than new ones are found.

    What it means for work. When answers get cheap, the scarce skill moves to choosing the right questions. That shift is coming to every field that runs on defined problems.

    Source: Terence Tao on Mathstodon, Sep 8, 2026.

  9. Adoption

    McKinsey: 39% expect AI to cut headcount next year

    In McKinsey's 2026 survey of 1,719 respondents, 39% expect AI to reduce their organization's total headcount over the next year, up from 32%. Among large enterprises, 40% are scaling AI agents, up from 27%.

    What it means for work. Workforce plans are being written around AI before the hiring data shows it. That is the gap this page exists to close.

    Source: McKinsey, The state of AI in 2026, Aug 25, 2026.

  10. Math and science

    AI systems report perfect scores at the 2026 Math Olympiad

    AFP reported that Huawei and Xiaohongshu said their systems scored 42 of 42 at IMO 2026. Xiaohongshu said its answers were submitted to the organisers for grading. A year earlier, Google DeepMind's gold-medal 35 of 42 was officially certified by IMO coordinators.

    What it means for work. Hard, well-posed problems with a checkable answer are now within reach of machines. Most paid work is less well-posed, which is where human judgment still earns its keep.

    Source: AFP via The Standard, Jul 23, 2026; Google DeepMind, Jul 21, 2025.

  11. Adoption

    One in five Canadian businesses now uses AI

    Statistics Canada reported that 19.2% of businesses used AI to produce goods or deliver services in the past 12 months, up from 6.1% two years earlier.

    What it means for work. Adoption roughly tripled in two years. The occupations inside those businesses feel it first in what the role requires, before it shows in postings counts.

    Source: Statistics Canada, Jun 11, 2026.

  12. Math and science

    OpenAI reports a model disproved Erdős's unit distance conjecture

    OpenAI said a model found a counterexample to Erdős's 1946 planar unit distance conjecture, checked by external mathematicians. Mathematician Daniel Litt called it the first AI result he found interesting in itself.

    What it means for work. AI moved from answering known questions to producing results experts value on their own merits.

    Source: Let's Data Science, May 28, 2026; Quanta Magazine, Aug 3, 2026.

  13. Usage

    Google processes 3.2 quadrillion tokens a month

    Google said it processes over 3.2 quadrillion tokens a month across its products, about seven times the 480 trillion of a year earlier.

    What it means for work. Use is compounding far faster than any hiring series moves. That is why use is a leading signal.

    Source: Google, Google I/O 2026, May 19, 2026.

  14. Capability

    Professional benchmark reaches 84.9%

    OpenAI reported that GPT-5.5 matched or beat an industry professional's deliverable in 84.9% of GDPval comparisons, up from 47.6% for the best model seven months earlier. OpenAI makes both the benchmark and the model.

    What it means for work. On defined, well-specified tasks, the best models now match or beat experienced professionals most of the time.

    Source: OpenAI, Introducing GPT-5.5, Apr 23, 2026.

  15. Capability

    Autonomous task length passes 16 hours

    METR measured Claude Mythos Preview (early) at a 50% time horizon of about 17 hours of skilled-human work, and noted its task suite cannot measure reliably above 16 hours. A year earlier the best measured model was at about 2 hours.

    What it means for work. Agents now finish, half the time, tasks that take a skilled person two working days. Work is being handed over in larger pieces.

    Source: METR, Time Horizon 1.1 results, entry added May 8, 2026.

  16. Usage

    Anthropic publishes AI use by occupation

    Anthropic's "observed exposure" estimates how much of each occupation's work time is on tasks where Claude is already used for work. Computer programmers scored highest at 75%. The authors found no systematic rise in unemployment for exposed workers, and suggestive evidence of slower hiring of 22 to 25 year olds into exposed occupations.

    What it means for work. This is the first public measure of actual AI use by occupation. JDI shows it on each occupation page next to hiring demand.

    Source: Anthropic, Labor market impacts of AI, Mar 5, 2026.

  17. Capability

    OpenAI launches GDPval

    GDPval scores AI against professionals on defined tasks from 44 occupations. The best model at launch, Claude Opus 4.1, matched or beat the professional in 47.6% of comparisons.

    What it means for work. The first benchmark built on the deliverables of real occupations, which makes it readable next to job data.

    Source: OpenAI, GDPval paper, Sep 25, 2025 (CC BY 4.0).

  18. Math and science

    AI earns an officially graded Math Olympiad gold

    Google DeepMind's Gemini Deep Think scored 35 of 42 at IMO 2025, graded and certified by IMO coordinators.

    What it means for work. The start of the run of math results above.

    Source: Google DeepMind, Jul 21, 2025.

What AI can do at work now

Three public measures, each drawn from the JDI record of publicly reported results. They measure different things on different scales, so they are kept apart.

Professional benchmark

Defined tasks from 44 occupations in the nine sectors that contribute most to US GDP, written by professionals with an average of 14 years of experience. Experts compare the AI's deliverable with a professional's, blind. GDPval, by OpenAI.

Working on its own

Multi-hour agentic tasks in 11 high-value fields, from chip design and hardware to finance, legal and software engineering. The AI works alone inside real software, with a default limit of eight hours per task. Headline score: pass@1 averaged over three attempts, each field weighted equally. GDPx v1.0, by SemiAnalysis.

Autonomous task length

The length of software and research tasks, measured by how long they take skilled humans, that an AI agent completes on its own half the time. Each point is the best model at its release date. Task-completion time horizon, by METR (Time Horizon 1.1).

The charts need JavaScript. Every figure is also in the dated record under Method and sources.

Since 2023 the task length has doubled about every 4.2 months on METR's estimate (129 days, with a range of 104 to 158). Growth like that stays small for a long time, then arrives all at once. The professional benchmark went from 47.6% to 84.9% in seven months. On GDPx, the newest and hardest of the three, the best publicly reported score is 16.4%: AI already finishes about one multi-hour professional task in six with no human help.

Professional benchmark, AI-judged rating

After April 2026, new GDPval results appear as Elo ratings from Artificial Analysis (GDPval-AA v2.1), judged by a panel of AI models. Elo is not a percentage and is not joined to the line above. We record the figures as public coverage reported them.

ModelEloReportedWhere we saw it
Claude Opus 5.5 (Max)1,866Oct 6, 2026Klu, GDPval-AA page quoting the Artificial Analysis board as of Oct 6, 2026
Claude Sonnet 5.5 (Max)1,839Oct 6, 2026Klu, as above
Claude Fable 5.2 (Max)1,831Nov 9, 2026BenchLeader, public mirror of the board, Nov 9, 2026

Use and spending: the other leading signal

Capability says what AI could do. Use shows what organizations already do with it: cut costs, speed up work and build the data centres to run it.

Use

3.2 quadrillion

tokens a month processed across Google's products and APIs, up from about 480 trillion a year earlier and 9.7 trillion in May 2024.

Google, Sundar Pichai, Google I/O 2026 keynote, May 19, 2026.

4×

Google's first-party model APIs process nearly 32 billion tokens a minute, more than four times the level a year earlier and 40% more than the previous quarter.

Alphabet, Q3 2026 earnings call, CEO remarks, Oct 28, 2026.

Spending on AI computing

$195–205B

Alphabet: 2026 capital expenditure guidance, unchanged; Alphabet expects a significant increase in 2027 and says it is buying third-party compute because supply is constrained.

Alphabet, Q3 2026 earnings call transcript, Oct 28, 2026.

about $220B

Amazon: capital spending CEO Andy Jassy expects for 2026, mostly for AI data centres, as reported by the Associated Press.

Associated Press, Oct 29, 2026.

$135–145B

Meta: 2026 capital expenditure outlook, including finance lease principal payments.

Meta, third quarter 2026 results, Oct 21, 2026.

$39.4B

Microsoft: additions to property and equipment in the July to September 2026 quarter, against $19.4B a year earlier.

Microsoft, FY27 first quarter results, Oct 28, 2026.

JDI's sum of the 2026 figures from Alphabet, Amazon and Meta, the three that give a full-year number. Microsoft reports by fiscal quarter and is left out of the sum. Together: $550B to $570B for 2026.

Adoption by employers

19.2%

of Canadian businesses used AI to produce goods or deliver services in the past 12 months, up from 12.2% a year earlier and 6.1% in the second quarter of 2024.

Statistics Canada, Analysis on AI use by businesses in Canada, Q2 2026 (Open Licence), Jun 11, 2026.

19.8%

of US businesses used AI in a business function in the survey period ending May 3, 2026. Among firms with 250 or more employees the share was 37%.

US Census Bureau, Business Trends and Outlook Survey (public domain), May 3, 2026.

39%

of respondents expect AI to reduce their organization's total headcount over the next year, up from 32% a year earlier. Among AI-using organizations, 14% say AI already contributed to a workforce decline.

McKinsey, The state of AI in 2026 (1,719 respondents, May 4 to June 8, 2026), Aug 25, 2026.

40%

of large-enterprise respondents are scaling AI agents in at least one function, up from 27% a year earlier.

McKinsey, The state of AI in 2026, Aug 25, 2026.

Employers are writing AI into job postings too. The share of postings that mention AI, by country and month, is on AI in job postings.

AI use by occupation, next to hiring demand

Anthropic Economic Index publishes an occupation-level measure called observed exposure: estimated share of an occupation's work time spent on tasks where Claude was already used for work in Anthropic's Aug and Nov 2025 samples. Automated use counts in full, assisted use at half weight. Tasks count only if a 2023 study judged them feasible for a language model. JDI matched it to 387 of the US occupations in the index by SOC code. Each occupation page now shows it.

The 12 occupations in the index with the highest observed AI use, and their US hiring demand
OccupationObserved AI useUS hiring demandSector postings, year on year
Computer programmers74.5%Mixed+23.4%
Medical records specialists66.7%Mixed−5.3%
Market research analysts and marketing specialists64.8%Rising+0.4%
Medical transcriptionists63.6%Falling−5.3%
Database architects57.9%Rising+10.0%
Financial and investment analysts57.2%Rising+11.7%
Software quality assurance analysts and testers51.9%Rising+23.4%
Statistical assistants51.0%Falling−2.0%
Information security analysts48.6%Rising+10.0%
Web developers48.0%Rising+23.4%
Mathematical science occupations, all other47.9%Rising+3.7%
Technical writers47.5%Steady+3.5%

Sources: Anthropic Economic Index, Labor market impacts of AI: A new measure and early evidence (Massenkoff and McCrory), Mar 5, 2026 (data released under CC-BY); JDI ratings and Indeed Hiring Lab sector postings, data to Nov 13, 2026.

High AI use has not yet meant falling demand

Observed AI useOccupationsRated RisingRated Cooling or Falling
25% or more6824 (35%)25 (37%)
10% to 25%7323 (32%)30 (41%)
Above 0, under 10%9840 (41%)28 (29%)
0% (no significant work use seen)14851 (34%)49 (33%)

JDI's count across the matched US occupations. So far the most-used occupations are about as likely to be rated Rising as Cooling or Falling. That is the lag at work: use shows up in what a role requires well before it shows up in how many people are hired. The postings figure is measured for a whole sector, so occupations in the same sector share it.

Limits: observed exposure counts Claude only, from samples in August and November 2025, and only tasks a 2023 study judged feasible for a language model. A low figure means Anthropic saw little work use of Claude on that occupation's tasks. It does not measure what AI can do in that occupation.

Disruption and physical AI

Capability and use are the leading signals. Disruption is what they do to particular occupations, and AI in machines moves on its own, slower clock.

Disruption Map

Every US occupation in the index, placed by its exposure to AI at a desk and on the floor against its hiring demand, in four zones, each with a next step. It also carries Disruption Watch and Opportunity Watch.

Physical AI

A dated, sourced record of humanoid and industrial robot milestones, from Figure, Tesla, Agility, Apptronik, 1X and Boston Dynamics to warehouse and factory deployments, with company claims reported as claims.

Method and sources

How the record is kept. JDI keeps its own dated record of AI capability results as they are publicly reported: lab announcements, the benchmark makers' own posts, and news coverage that reads the figures. Each row carries the date, the model, the figure, its exact wording and where we saw it. We do not scrape the Artificial Analysis or GDPx sites, and we have not asked either for permission; figures from them are recorded as public coverage reported them.

Versions stay apart. GDPval wins-or-ties percentages, GDPval-AA Elo ratings, GDPx scores and METR time horizons are different measures. None is converted into another. When a benchmark changes version, the line stops and the new version starts a new series.

Who reported it. OpenAI makes GDPval and sells models scored on it, so GDPval figures from OpenAI launch posts are vendor-reported. The METR figures come from METR's published results file; METR says its page is no longer actively updated (Sep 8, 2026) and that measurements above 16 hours are unreliable with its current tasks. Usage and spending figures are the companies' own statements. Adoption figures are from Statistics Canada, the US Census Bureau and McKinsey's survey.

What these measures leave out. Every benchmark here scores defined tasks with a checkable result. None scores judgment about what matters to a business, accountability for an outcome, or the trust a decision maker places in a person. That is the blind spot of AI, and it is why these scores are read beside hiring demand rather than in place of it.

Not covered. Company share prices, chip-maker earnings and executive news are left out on purpose.

The dated record

DateMeasureModelResultWhere we saw it
Nov 9, 2026Professional benchmark, AI-judged rating (GDPval-AA v2.1, Elo)Claude Fable 5.2 (Max)1,831BenchLeader, public mirror of the board, Nov 9, 2026
Nov 9, 2026Working on its own (GDPx v1.0)Claude Fable 5.216.4%ArtificialWatch, Nov 10, 2026, reading the GDPx leaderboard
Nov 9, 2026Working on its own (GDPx v1.0)GPT-6.1 Terra8.5%BenchLeader, public mirror, Nov 9, 2026
Nov 9, 2026Working on its own (GDPx v1.0)GPT-6.1 Sol8.4%BenchLeader, public mirror, Nov 9, 2026
Nov 2, 2026Working on its own (GDPx v1.0)Claude Opus 5.512.0%Unite.AI, Nov 2, 2026, on the SemiAnalysis launch
Nov 2, 2026Working on its own (GDPx v1.0)Gemini 4 Carbon10.9%Unite.AI, Nov 2, 2026
Nov 2, 2026Working on its own (GDPx v1.0)GPT-6 Astra9.6%Unite.AI, Nov 2, 2026
Nov 2, 2026Working on its own (GDPx v1.0)Muse Wonder Full9.0%Unite.AI, Nov 2, 2026
Nov 2, 2026Working on its own (GDPx v1.0)Claude Fable 5.18.5%DataStudios, Nov 2026, full launch table
Oct 6, 2026Professional benchmark, AI-judged rating (GDPval-AA v2.1, Elo)Claude Opus 5.5 (Max)1,866Klu, GDPval-AA page quoting the Artificial Analysis board as of Oct 6, 2026
Oct 6, 2026Professional benchmark, AI-judged rating (GDPval-AA v2.1, Elo)Claude Sonnet 5.5 (Max)1,839Klu, as above
Apr 23, 2026Professional benchmark (GDPval, wins or ties)GPT-5.584.9%OpenAI, Introducing GPT-5.5, Apr 23, 2026
Apr 7, 2026Autonomous task length (METR, 50% horizon)Claude Mythos Preview (early)17.4 hMETR, Time Horizon 1.1 benchmark results file
Mar 5, 2026Professional benchmark (GDPval, wins or ties)GPT-5.483.0%OpenAI, Introducing GPT-5.4, Mar 5, 2026
Feb 5, 2026Autonomous task length (METR, 50% horizon)Claude Opus 4.612.0 hMETR, Time Horizon 1.1 benchmark results file
Dec 11, 2025Professional benchmark (GDPval, wins or ties)GPT-5.2 Pro74.1%OpenAI, Introducing GPT-5.2, Dec 11, 2025
Dec 11, 2025Autonomous task length (METR, 50% horizon)GPT-5.25.9 hMETR, Time Horizon 1.1 benchmark results file
Nov 24, 2025Autonomous task length (METR, 50% horizon)Claude Opus 4.54.9 hMETR, Time Horizon 1.1 benchmark results file
Nov 18, 2025Autonomous task length (METR, 50% horizon)Gemini 3 Pro3.7 hMETR, Time Horizon 1.1 benchmark results file
Sep 25, 2025Professional benchmark (GDPval, wins or ties)Claude Opus 4.147.6%OpenAI, GDPval paper (CC BY 4.0), Sep 25, 2025
Aug 7, 2025Autonomous task length (METR, 50% horizon)GPT-53.4 hMETR, Time Horizon 1.1 benchmark results file
Apr 16, 2025Autonomous task length (METR, 50% horizon)o32.0 hMETR, Time Horizon 1.1 benchmark results file
Feb 24, 2025Autonomous task length (METR, 50% horizon)Claude 3.7 Sonnet1.0 hMETR, Time Horizon 1.1 benchmark results file
Dec 5, 2024Autonomous task length (METR, 50% horizon)o139 minMETR, Time Horizon 1.1 benchmark results file
Oct 22, 2024Autonomous task length (METR, 50% horizon)Claude 3.5 Sonnet (new)20 minMETR, Time Horizon 1.1 benchmark results file
Sep 12, 2024Autonomous task length (METR, 50% horizon)o1-preview20 minMETR, Time Horizon 1.1 benchmark results file
Jun 20, 2024Autonomous task length (METR, 50% horizon)Claude 3.5 Sonnet11 minMETR, Time Horizon 1.1 benchmark results file
May 13, 2024Autonomous task length (METR, 50% horizon)GPT-4o7 minMETR, Time Horizon 1.1 benchmark results file
Nov 6, 2023Autonomous task length (METR, 50% horizon)GPT-4 Turbo4 minMETR, Time Horizon 1.1 benchmark results file
Mar 14, 2023Autonomous task length (METR, 50% horizon)GPT-44 minMETR, Time Horizon 1.1 benchmark results file

Data files: capability.json, demand.json, log.json, aei_exposure.json. Retrieved Nov 17, 2026. How the hiring ratings are built: method. Written analysis: AI and jobs insights.

Sources

These links open in a new tab.