What AI can do at work now: the professional benchmarks, explained https://jobsdemandindex.com/insights/what-ai-can-do-at-work-professional-benchmarks Jobs Demand Index. Published Nov 17, 2026. Data to Oct 30, 2026. Topic: AI and jobs. Tags: GDPval, GDPx, METR time horizon, AI benchmarks for jobs, can AI do my job, AI agents at work. Summary: On defined tasks the best AI matches or beats professionals 84.9% of the time (GDPval). Working alone for hours, the best model completes 16.4% of tasks (GDPx). Key facts - OpenAI reported on Apr 23, 2026 that GPT-5.5 matched or beat an industry professional's deliverable in 84.9% of GDPval comparisons; the best model at GDPval's launch on Sep 25, 2025 scored 47.6%. - SemiAnalysis launched GDPx v1.0 on Nov 2, 2026, with multi-hour tasks in 11 fields; the best model scored about 12% at launch, and Claude Fable 5.2 was reported at 16.4% on Nov 9, 2026. - METR measured the best AI agent's 50% task-completion horizon at about 17 hours of skilled-human work in April 2026, with a doubling time since 2023 of about 129 days. - None of the three benchmarks scores judgment about business value, accountability for outcomes or the trust of decision makers. "Can AI do my job?" is the wrong question for most professionals. AI is already doing part of most knowledge jobs, and the part is growing. Three public benchmarks measure how large that part is, and each measures something different. This guide explains them as of Nov 17, 2026, beside the index data to Oct 30, 2026. Three measures, three scales | Benchmark | What it tests | Latest public result | |---|---|---| | GDPval (OpenAI), our "professional benchmark" | Defined tasks from 44 occupations; experts compare the AI's deliverable with a professional's, blind | 84.9%, GPT-5.5, Apr 23, 2026 (vendor-reported) | | GDPx v1.0 (SemiAnalysis), our "working on its own" | Multi-hour tasks in 11 fields, done alone inside real software, up to eight hours each | 16.4%, Claude Fable 5.2, reported Nov 9, 2026 | | Time horizon (METR) | The length of task, in skilled-human time, an agent finishes alone half the time | About 17 hours, Claude Mythos Preview, Apr 2026 | The scales are not interchangeable. GDPval asks whether a finished deliverable is as good as a professional's. GDPx asks whether the AI can get to a finished result with nobody helping. METR asks how long a task can be before the agent fails half the time. How fast they are moving GDPval went from 47.6% to 84.9% in seven months. OpenAI makes the benchmark and sells the model, so treat its launch-post figures as vendor-reported. On METR's estimate the task horizon has doubled about every 129 days since 2023. GDPx is new and hard: under strict pass-or-fail grading every model scored zero on finance and legal at launch. Still, about one multi-hour professional task in six now gets done with no human help. Six occupations covered by GDPval tasks, with US hiring demand: - Software developers: US hiring demand Rising, US median pay $135,980 (https://jobsdemandindex.com/occupation/15-1252/) - Accountants and auditors: US hiring demand Cooling, US median pay $83,680 (https://jobsdemandindex.com/occupation/13-2011/) - Financial and investment analysts: US hiring demand Rising, US median pay $102,740 (https://jobsdemandindex.com/occupation/13-2051/) - Lawyers: US hiring demand Steady, US median pay $159,670 (https://jobsdemandindex.com/occupation/23-1011/) - Mechanical engineers: US hiring demand Rising, US median pay $104,110 (https://jobsdemandindex.com/occupation/17-2141/) - Computer and information systems managers: US hiring demand Rising, US median pay $175,140 (https://jobsdemandindex.com/occupation/11-3021/) Every occupation in the table is covered by GDPval tasks, and most are rated Rising in the US. High benchmark scores and steady hiring sit side by side for now. What the benchmarks leave out Each test scores a defined task with a checkable result. None scores the work around it: deciding which problem is worth solving, owning the outcome when it goes wrong, and earning the trust of the decision maker who signs off. That is the blind spot of AI, and it is where a professional's leverage is moving. What to do with this Look up your occupation in the occupations search (https://jobsdemandindex.com/map) and read its hiring demand next to its observed AI use. Every benchmark result we track, with dates and sources, is on the AI Frontier page (https://jobsdemandindex.com/frontier). Limits: benchmark tasks are samples, not whole jobs. Results come from public coverage and lab announcements, recorded with the date we saw them. Sources - OpenAI, GDPval paper, Sep 25, 2025, CC BY 4.0 (arxiv.org, 2510.04374) - OpenAI, Introducing GPT-5.5, Apr 23, 2026 (openai.com) - SemiAnalysis, GDPx: measuring frontier model capabilities, Nov 2, 2026 (newsletter.semianalysis.com) - ArtificialWatch, Claude Fable 5.2 tops SemiAnalysis GDPx at 16.4 percent, Nov 10, 2026 (artificialwatch.com) - METR, Time Horizon 1.1 results file, retrieved Nov 17, 2026 (metr.org) Cite: Jobs Demand Index, "What AI can do at work now: the professional benchmarks, explained", Nov 17, 2026 (data to Oct 30, 2026). https://jobsdemandindex.com/insights/what-ai-can-do-at-work-professional-benchmarks Reuse: quote and cite with attribution and a link. Terms: https://jobsdemandindex.com/data