Human professional
—
Inside the price
The same task, done by a human mind or a machine one — priced in what each charges, and what each burns.
loading…
Pick a task. Both sides show the market price of the completed task — what a buyer actually pays — then the same task priced in energy and water. Open each invoice to see what it's made of.
Quality parity assumption: both invoices assume the output is accepted at comparable quality. That increasingly holds for routine versions of these tasks, and does not yet hold for the hardest versions — where the honest AI price includes human review time.
Explore each lens in depth: price over time · energy · water · your live grid · the policy tool
Every hard problem humanity faces is, at bottom, a problem of applying intelligence. Curing disease, decarbonising the grid, teaching children… just about any challenge you can think of is bounded by how much skilled thinking we can afford to point at it. For all of history, that thinking came from exactly one source: human minds — expensive to train, expensive to sustain, and impossible to copy.
That constraint is breaking. For the first time, cognition is something you can buy by the token, and its price is collapsing — GPT-4-level capability costs roughly 300× less than it did in March 2023. When a fundamental input gets cheap this fast, history says everything downstream reorganises. Artificial light once cost a day's wages for an hour of it; when it became too cheap to notice, it changed what a day is. Computation and communication told the same story, each time remaking work, cities, and lives.
Cheap intelligence is a bigger version of that story, because intelligence is the input that produces every other input. A tutor for every child, a second medical opinion for every scan, an engineer's attention for every village water pump. These things were never uneconomic because they lacked value. They were uneconomic because intelligence was scarce. The bridge between today and a radically better future is intelligence and its application.
You can already watch this happening in the data. The frontier labs now publish task-level maps of their own economic footprint — Anthropic's Economic Index and Google's ATLAS, which between them track thousands of tasks across hundreds of occupations. Two findings recur: today's AI augments far more than it replaces — roughly 57% of use versus 43% — and adoption follows income, compounding the lead of places already ahead. Those are the same tasks this site prices.
But transitions this large are not automatically beneficial. They move wages, remake professions, concentrate capability, and draw real megawatts and real water from real places. In short, the radical changes in the digital world have real world consequences. Steering well requires seeing clearly, and much of the public conversation runs on vibes in both directions.
Nowhere is the choice sharper than in Britain. Seventy-three per cent of the British workforce now uses AI — up from a third a year earlier — and Google alone reckons its tools already add £140 billion to UK activity, most of it flowing through small firms. Britain founded the modern AI lab; its grid — often a third wind on a good day — is clean enough that compute here is low-carbon almost by default. The raw materials of an AI-led growth decade are already on the island.
And yet Britain keeps tying its own shoelaces together. It charges some of the rich world's highest industrial electricity prices, waits years to approve the data centres that sovereign AI depends on, and deploys AI most timidly in exactly the public services where the gains would be largest. The binding constraint is not the technology, and — as the live grid and water pages show — it is not the physics. It is nerve. The narrower policy question, should this particular deployment go ahead?, is what GreenBench is for.
Cheap intelligence is the clearest growth lever Britain has left. The only question is whether it pulls it — or prices, permits and regulates its way out of the one advantage it still holds.
List price to draft a 1,000-word article, at each era's model prices. The frontier flagship stays premium; the price of yesterday's frontier collapses — GPT-4-level capability costs roughly 300× less than it did in March 2023.
USD per task · archived launch list prices; hollow points are estimates
A human runs on roughly 2,000 kcal a day — about 2.3 kWh of energy, bought as food. A model runs on grid electricity. The fuels differ in price per unit as much as the workers differ in price per task.
USD per kWh delivered to the worker · US averages · see your grid live →
Watt-hours per completed task · every gridline labelled, one order of magnitude past the data
The rebound: per-task energy is tiny, but total AI electricity demand is growing fast even as each task gets leaner. This is often called Jevons' paradox — though it is less a paradox than a predictable release of demand once AI sharply lowers the private marginal cost of a service. Per-task figures answer “what does this cost?”, not “what will the world use?” — for the deployment-level question, see GreenBench.
Water is in the ledger above; its full accounting — on-site cooling vs the full chain, closed-loop cooling, and why siting matters — lives on its own page. So does the live grid, which reprices AI energy by your region and the current wind.
Educating one human costs far less than training one frontier model. But a model is copied and shared across trillions of tasks, while a human mind serves one career. Amortisation reverses the comparison.
Public spending on one bachelor's-track education, over ~17 years.
Representative frontier training run, 2025–26 ($200–500M range).
Per-task numbers answer “what does this cost?”. Governing AI needs a different question: should this deployment happen at all? GreenBench turns the measurement on this site into a three-gate permission standard.
Run a deployment through three gates: integrity (is the ecological burden auditable?), counterfactual justification (does it beat human work, conventional software, and no action — retries and oversight included?), and justice (can this place carry it?). The verdict updates as you answer: approve, approve with conditions, or refuse.
For each worker, cost is examined through three layers:
These layers are lenses, not addends. A wage already repays groceries and tuition; an API price already covers the power bill and amortises 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.
Businesses and governments do not buy prompts; they buy services — completed jobs at an agreed quality. When output quality is poor, retries, human checking, and downstream correction expand the true labour and compute cost while disappearing from headline per-prompt figures. That is why every number on this site is priced per completed task, why the ledger carries a hybrid figure with human review time included, and why GreenBench benchmarks deployments against a human-work baseline, a conventional-software baseline, and a no-action baseline. It is also how the frontier labs now measure themselves: Anthropic's Economic Index and Google's ATLAS (≈4,000 tasks, 800 occupations) both map AI usage by task and occupation, not by prompt.
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.Every chart above, as numbers (current model selection applies):