Anthropic's Economics team published a page, "What will our economic future look like?" Under it is a working paper by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory: Economic Scenarios for Transformative AI, September 2026. You set how capable AI will be, how widely it is used, how much it does alone, how much more productive it makes people, and how long a new job takes. The model returns a 2030 United States.
The economy, here, is bundles of tasks. A nurse's day is a list from O*NET. AI can leave a task alone, help a person do it, do it without the person, or create a new one. Add up every instance and you get GDP. The authors attach no probabilities. Modest, substantial, and extreme are three settings of the same knobs.
Modest: GDP 1.6 percent above a path with no AI, about $34.1 trillion at 2025 prices. Unemployment up a tenth of a point. They liken it to the internet. Substantial: GDP 8.3 percent higher, about $36.3 trillion. Knowledge wages roughly flat. Other wages up. They surveyed 10,980 US adults; the median answers land close to this, with GDP about 8 percent higher and cognitive employment down about 4 percent. Extreme: GDP 32 percent higher, about $44.4 trillion, with growth around 15 percent a year. AI performs almost half of today's cognitive work. Labor's share of each dollar falls from about 60 cents to 45. Knowledge wages 11.5 percent below the no-AI path. Other wages about 34 percent above it. Nearly one in five cognitive workers unemployed.
Why capital takes more by 2030 is in the paper's own nutshell. In the long run, they write, new machines can always be built, so productivity gains accrue to labor. Between now and 2030, capital is slower to adjust, so some of the gain shows up as a higher return to owners. There are only two occupation groups, cognitive and everyone else. A person who switches earns the other group's wage at once, with no discount for lost tenure. Recursive self-improvement sits in the extreme bucket because it is written into the capability path. The ideas block is a Jones-style research engine bottlenecked by physical tasks. On the extreme path, the ideas stock is 0.6 percent above its no-AI path by 2030. The model cannot generate an explosive research takeoff. They stop at 2030 partly because they do not let robotics move physical work.
Editorial
The labor-share fall is not a law of AI. It is what their 2030 window does when capital cannot fully adjust. In their long-run nutshell, gains accrue to labor. They do not take the model past 2030. Keep the window short and owners take more of a larger pie even if average pay rises. That is the result I keep.
Recursive self-improvement is written into the capability path because the research block cannot discover it. Research stays tied to physical tasks, so the ideas stock barely moves by 2030 even in extreme. The lab that may build that path is also the lab publishing the slider. Their survey maps the public onto substantial.
They are honest that today's quiet labor market has two readings: modest forever, or the first inch of extreme. Those readings ask for different institutions. Jacob Coxon left Anthropic over a race to self-improving systems. In this model that race is the extreme capability path, not a finding. The page will not tell you which reading we are in.