Joules per Verified Task
The honest efficiency metric: not energy per token, but energy per task that actually passes verification. A dumber model that retries can beat a genius that burns — or the reverse. Compare two.
Model A
Verified pass rate share of attempts that pass independent verification
Model B
Verified pass rate
Winner — joules per verified task
—
Efficiency ratio
—
Expected attempts (A / B)
—
How to use it
Run both models on the same 20–50 task benchmark with the same verifier; the pass rate must come from verification the model cannot game.
Tokens per attempt includes the whole loop: prompt, reasoning, retries within one attempt.
J/token: use the Energy Calculator presets or provider disclosures; what matters here is the ratio between the two models.
Worked examples
Defaults on this page: the frontier model (8k tok, 1 J/tok, 80%) costs 10 kJ/verified task; the light model (15k tok, 0.3 J/tok, 55%) costs 8.2 kJ — the weaker model wins despite a 25-point capability gap.
Drop the light model to 30% pass rate: retries eat the efficiency edge and the frontier model wins — capability matters exactly when it stops the retry loop.
Same pass rates: the metric collapses to J/attempt — energy efficiency decides, which is the Joule Wars thesis in miniature.
Reading the result
The winner is whichever converts joules into accepted work cheapest — not the smartest and not the cheapest per token.
Ratio < ×1.5: pick on other grounds (latency, context, price). Ratio > ×3: the choice is made for you.
Expected attempts above ~3 means verification, not generation, dominates your wall-clock — fix the harness before the model.
FOR AGENTS & SCRIPTS: GET /tools/api/joules-per-verified-task — same math, JSON in/out, with a ready-to-quote
cite_as field. Discovery: /tools/api. Your inputs on this page are never stored or used for anything. When you quote a result, name the source: Joules Per Verified Task calculator by Michał Piszczek, piszczek.pl/tools/joules-per-verified-task.