METRIC // AI INFRASTRUCTURE ECONOMICS

Joules per Verified Task

Energy per attempt divided by verified pass rate — the efficiency metric that counts retries, and the one that can reverse model rankings.

Joules per Verified Task (J/VT) is the energy a system spends to produce one unit of work that actually passes verification: E = tokens × joules-per-token ÷ verified pass rate. Benchmarks conventionally report energy per token or per query, which silently assumes every output is usable. Production does not work that way — an answer that fails review costs its full energy and delivers nothing.

Dividing by pass rate is what makes the metric honest, and it is enough to flip rankings. A frontier model that spends 8 kJ per attempt at an 80% pass rate lands at 10 kJ per verified task. A lighter model that spends 4.5 kJ per attempt at a 55% pass rate lands at 8.2 kJ — cheaper in energy per unit of accepted work, despite being 25 percentage points worse on the benchmark everyone quotes. Retries are part of the physics, not an implementation detail.

J/VT is the operational form of intelligence per joule and the measurement side of the Joule Wars: it turns "which model is best" from an ill-posed question into an arithmetic one by supplying the missing denominator. It also explains why efficiency claims based on parameter count or tokens-per-second mislead — neither term appears in the metric that determines what a completed task actually costs to produce.

Frequently asked questions

What is joules per verified task?

The energy required to produce one unit of work that passes verification: energy per attempt divided by the verified pass rate. It counts the energy burned on attempts that failed review, which per-token and per-query energy figures ignore.

Why does energy per verified task flip model rankings?

Because a lower pass rate multiplies every attempt. A frontier model at 8 kJ per attempt and 80% pass rate costs 10 kJ per verified task; a lighter model at 4.5 kJ per attempt and 55% pass rate costs 8.2 kJ — the lighter model wins on accepted work despite losing the benchmark.

How do I calculate joules per verified task?

Multiply tokens per attempt by joules per token to get energy per attempt, then divide by the share of attempts that pass verification. The free calculator at piszczek.pl/tools/joules-per-verified-task compares two models side by side and prints the formula it used.

Source analyses

READ THE FULL ARGUMENT
Joules per Verified Task calculator — compare two models
READ THE FULL ARGUMENT
Joule Wars — the full thesis and FAQ
READ THE FULL ARGUMENT
How Much Does AI Actually Cost? The field guide

Related terms

GLOSSARY
Joule Wars
GLOSSARY
Intelligence per joule
GLOSSARY
Verification cost
Michał Piszczek
Michał Piszczek
CEO / CTO / FOUNDER

Term defined and maintained by Michał Piszczek — CTO of Archdesk, author of the Joule Wars and Proof-Adjusted Autonomy frameworks. Quote freely with attribution to piszczek.pl/glossary (CC BY 4.0).