Elliott CheeksCost-Per-Outcome, Not Cost-Per-Token
The token price is the smallest number in your AI economics; the one that matters is what it costs to produce an output that survives verification.
Every AI pricing conversation I hear starts in the wrong place.
Teams compare models on cost per million tokens the way they compare fuel prices per gallon. Model A is $3, Model B is $15, so Model A is the responsible choice.
Procurement nods. The pilot gets approved on the cheap model. Everyone feels disciplined.
Then the outputs start arriving, and nobody is measuring the number that actually hits the P&L.
The unit of value is not a token
A token is not a deliverable.
Nobody’s business runs on tokens. It runs on outcomes: a commission statement that pays people correctly, a report an executive acts on, a forecast a trader positions against, a price curve that does not crash the overnight run.
The honest unit of AI cost is cost per verified outcome:
What it costs to produce one output that survives contact with a known number.
The formula is simple:
Cost per outcome = (model spend + review time + rework + cost of being wrong) / outputs that pass verification
Look at what is in the numerator.
Token spend is one term. In production systems, it is often the smallest one.
A worked example
The numbers below are illustrative, but the structure is from real life.
Say you have a monthly reconciliation report. Two paths.
Path A uses the cheap model and ships without a gate.
I ship AI-built software into production and write about the gates, patterns, and cost discipline that make it work. More about me →
Practitioner notes on the gates, the patterns, and the cost discipline behind AI-built software.
Related reading
Seven gates, each one blocking the next. Most projects should die at Gate 1, and the ones that survive all seven are the ones people actually trust.
Prompting, examples, review: you learned all of it the first time you delegated. AI fluency is management fluency with a shorter feedback loop.