AI and the US economy: move the parameters yourself

An independent reimplementation of the Anthropic Institute's 2026 model, exposing the parameters their own explorer holds fixed

The Anthropic Institute's scenario explorer lets you move the five parameters describing how far AI gets and how fast. This one adds the three it holds constant, which is where the interesting questions live: how easily the economy can add capital, how readily different kinds of work substitute, and how much AI use leaves the paid economy altogether.

Everything runs in your browser. Nothing is sent anywhere.

Scenarios

Start from one of the paper's three, then change anything.

How far AI gets their five

Share of tasks AI can do

m, capped at all cognitive work (0.624)

Share of those actually using it

d, diffusion

Cost saving per task

a, in log points

Of AI use, share that replaces a worker

ψ, automation rather than augmentation

New tasks created per task automated

ρ, reinstatement

What they hold fixed the point of this page

How easily capital can be added

ε. Their assumption is 3.0. Below the threshold shown on the right, the average wage falls.

How readily work substitutes

σ. Their assumption is 0.5. Watch how little the labour share moves.

Share of AI use that leaves the market

h. Not in their model at all. Work you stop paying for and do yourself.

Labour market my reproduction

Normal quit rate, per year

How fast growing sectors recruit

θ_H, share of the shortfall posted monthly

Ease of switching occupation

μ, search discount when crossing

Wage stickiness

ξ. Off by default: see the caveat below.

Use the sticky-wage channel

Turning this on makes my fit to their published figures worse, not better. Left off for that reason.

Results in 2030

quantity	this model	their published

green quantities come from the closed-form block, which reproduces their published figures closely. amber ones come from my labour-market reproduction, which does not match theirs exactly. The "their published" column is fixed at the paper's numbers for the nearest scenario and will not track your slider changes.

The wage question

Measured against actual

Self-test

This page recomputes the paper's three scenarios on load and compares against their published figures. If the numbers below drift, the page is wrong and you should not trust the rest of it.

running...

What this is, and what it is not

This is an independent reimplementation, written from the equations printed in the paper. The authors publish no code, so nothing here is derived from theirs. The closed-form block reproduces their published output figures to within about 0.35 percentage points once the two components I do not model are accounted for. The labour-market block reproduces unemployment to within about 0.6 points, and does not match their implementation; its outputs are labelled accordingly.

Parts of the underlying model are formalised in Lean 4, a proof assistant, including the theorem that the model has exactly one solution and the arithmetic behind the measurement panel. Those checks apply to the mathematics, not to this JavaScript, which is an ordinary port and carries no such guarantee.

I do not model the innovation block, which moves the published results by under 0.61 points in any scenario, or the price rigidity and demand feedback the authors also exclude. Their caveats section is worth reading and is unusually candid.

Background and the findings this was built to make touchable: The number Anthropic's economists left as a word.