Live tracker · data updated Oct 19, 2026
Two clocks, side by side. The fast one tracks what AI can do and what it costs, synced daily. The slow one tracks productivity and jobs in eight economies. The gap between them is the story, and the weak-link view of growth explains it.
Come back any time
This page remembers when you last visited (in your browser only) and shows what changed: new models, capability records, price drops, and new labor data.
Latest: Meta AI released Muse Spark OpenOct 15, 2026
The fast clock · global
650× cheaper
GPT-4-level intelligence cost $37.50 per million tokens in Mar 2023. Today 135 models from 14 labs match it, 55 of them open-weight, from $0.058. See every level
Source: Capability: Epoch Capabilities Index. Prices: providers' list prices, blended 3:1 input to output, per token. Task horizons: METR. · Epoch Capabilities Index (ECI) · Data on AI models · LiteLLM model price map · Task-completion time horizons (Time Horizon 1.1)
The slow clock · by country
OECD average
Context, not cause: unemployment moves with interest rates and the business cycle, and productivity with everything else. Output per hour comes from the BLS (US, quarterly); output per worker from the World Bank (annual, comparable across countries).
Source: Unemployment: BLS (US), OECD harmonised rates, Eurostat (EU). Productivity: BLS output per hour (US) and World Bank output per worker. · US Bureau of Labor Statistics · OECD Data Explorer · Eurostat · World Development Indicators
Answer six questions and see which economist's view you're closest to.
Take the quiz →See where people use AI across your occupation's tasks.
Find your occupation →Find what's stopping AI from making your workflow faster.
Map your weak links →Since ChatGPT launched (Nov 2022), the price of GPT-4-level intelligence has fallen 650× and the length of task a frontier model can finish has grown 1740×. Over the same stretch, US output per hour rose 9.1% in total. Cheap inputs do not finish the job: work is a chain of tasks, and output waits on the links still held by people, such as judgment, accountability, trust, and physical presence.
Read the explainerFrom capability to value
The gap between the clocks is not just a matter of waiting. It also depends on the applications we build: do they take over existing tasks, help people do better work, or both?
Reduce the human effort needed for an activity. But producing a draft is not the same as delivering an accepted result: review, exceptions, and accountability still count.
Illustrative example: extract fields from invoices, with people handling unusual cases and checking errors.
Measure: total time and cost per accepted result, including rework.
Help a person make better decisions, serve more people, or offer something that was previously impractical. Value can come from better output, not just fewer hours.
Illustrative example: help a support agent find relevant information while the agent handles the customer and decides what to do.
Measure: quality, capacity, and who benefits, as well as time saved.
These routes overlap within a job. Neither an automation label nor an augmentation label tells us whether employment or wages will rise or fall; demand, adoption, and how the gains are shared matter too. The mechanisms apply across countries, but outcomes need local evidence.
Our price comparison holds benchmark performance roughly fixed. It does not measure the full cost of a workflow: integration, verification, rework, and human responsibility remain. Faster models are inputs; useful applications turn them into outcomes.
Source: An interpretation of the automation/complementarity distinction, not a replacement forecast. Acemoglu argues for a more worker-complementing direction; this is a perspective, not a settled outcome. · Acemoglu, Humanist Review · The Simple Macroeconomics of AI
Explore
Pick a capability level and see how many models reach it today, from how many labs, and how much cheaper it has become since it debuted.
Every notable release, its independent capability score, price, and lab country, and how fast open-weight models catch the frontier.
Is AI showing up in jobs yet? Postings in exposed versus hands-on work, youth unemployment, and firm adoption across eight economies.
Search any of about 900 occupations, see where people already use AI for each task, and find the weak links that keep a person needed.
When could AI finish a task as long as a workday, a week, or a year? Extend METR's measured trend and see the range.
What CEOs, researchers, and forecasters said about AI, the date they named, and what happened. Too fast, too slow, or still open.
Six plain-language questions about AI and work. See which economist's forecast yours is closest to.
Why isn't AI making your team faster? Map a workflow, have the people who do it rate each step, and see what to automate and what to redesign.
Why fast capability meets slow growth: weak links, the 150-year 2% line, the Jones & Tonetti forecast, and the strongest objections.