The Price of Intelligence

The framework: three layers, one price

For each worker, cost is examined through three layers:

1. Marginal — the market price of the completed task (a freelance rate, a wage × time, an API bill).
2. Sustaining — the fuel that keeps the worker running: food for the human, electricity for the model.
3. Training — the capital that made the worker capable: education for the human, the pre-training run for the model.

These layers are lenses, not addends. A wage already repays groceries and tuition; an API price already covers the power bill and amortizes the training run. So the headline comparison is always market price vs. market price, and the “inside the price” bars show how much of each invoice its fuel and training actually account for. The striking result: for both workers, fuel is a rounding error — the price of intelligence is almost entirely capital and scarcity, not energy.

Data confidence and freshness

• Every model price carries a confidence badge. Confirmed means the vendor's own list price; reported means a marketplace or third-party figure (open-weight models have no single vendor price); estimated means a representative number where sources conflict or vary by host; unavailable means no honest single number exists, and the site shows “—” rather than inventing one.
• All data lives in dated snapshots (data/YYYY-MM-DD.json). The “Data as of” picker in the top bar switches snapshots; a refresh script (scripts/update-data.mjs) pulls current model prices from the OpenRouter API and official statistics where keys are available, and appends a new snapshot — so the site accumulates its own history as sources update.

Key conventions

• Human energy accounting: a worker's full 2,000 kcal/day (2.33 kWh) is allocated across an 8-hour workday → 291 Wh per working hour. This mirrors the AI side, where Google's published per-prompt figure includes idle machines and datacenter overhead — both sides carry their standby costs.
• AI energy scaling: anchored to published medians (0.24 Wh per median Gemini prompt, 0.34 Wh per average ChatGPT query) and scaled linearly with tokens processed. This is a modelling assumption, not a measurement; heavy reasoning tasks land in the tens of watt-hours, consistent with press estimates for reasoning-mode queries.
• AI fuel price: industrial electricity (8.66¢/kWh), the closest EIA class to datacenter rates. The human's food is priced at the USDA moderate plan (~$15.50/day).
• Human training: 13 years of K-12 at $16,080/yr plus 4 years of public college at $36,101/yr ≈ $353k, amortized over a 40-year, 72,000-hour career.
• AI training: a $300M representative frontier run amortized over ~1 trillion lifetime queries, scaled per task by token volume. Training electricity is estimated from Epoch AI's finding that energy is ~2–6% of development cost.

What this comparison is not

• Not a claim of interchangeability. Tasks are not jobs. The human price bundles accountability, context, relationships, and liability that an API call does not carry.
• Not a physics claim about wages. Human labor is expensive because human time is scarce and lives are expensive to sustain in full — not because metabolic energy costs $6.66/kWh.
• Not stable. AI prices per unit of capability have fallen sharply year over year — see section 03 — and each snapshot is dated for exactly that reason.
• US-centric. Wages, food, electricity, and tuition all vary enormously by country. The live page localizes the energy lens where free data allows.

The full table

Every chart above, as numbers (current model selection applies):

Sources