Cost per outcome misses the target

By Alan Jacobson, AI Economics Strategist

Digital pricing has always followed a simple pattern: move closer to the result, and pricing gets smarter.

CPM (cost per thousand impressions): pay for exposure
CPC (cost per click): pay for engagement
CPA (cost per action): pay for a result

Each step reduced waste because the system could clearly observe what happened:

An impression is visible
A click is discrete
A conversion is verifiable

These models worked because the underlying system was simple, observable, and linear.

The unit being priced (impression, click, action) was directly measurable and tightly coupled to the event.

So the instinct to move one step further — to Cost Per Outcome (CPO) — feels logical.

If we can price an action, why not price the outcome itself?

Why people are proposing cost per outcome

Because tokens don’t work.

Tokens are currently used as the unit of measurement in AI systems, but they fail in two fundamental ways:

Semantically blind: they treat all words as if they carry equal cognitive and computational weight
Operationally blind: they miss large portions of agentic execution — tool calls, retrieval, loops, background tasks — because those processes don’t generate tokens

So what happens?

You get a system where:

usage looks stable
costs fluctuate underneath
and no one can explain why

When the unit of measurement fails, people don’t fix the unit.
They move up the stack.

That’s where “cost per outcome” comes from.

It’s an attempt to bypass a broken measurement layer entirely:

“If we can’t measure the work, let’s just price the result.”

On the surface, that sounds elegant.

In practice, it breaks down immediately.

Why cost per outcome won’t work

Cost per outcome assumes something that isn’t true: that outcomes are stable, observable, and attributable.

They’re not.

1. Outcomes are not discrete events

A “conversion” is clear.
An “outcome” in AI is not.

Is the outcome the answer?
The quality of reasoning?
The downstream business impact?
The absence of error?

Unlike CPM, CPC, or CPA, outcomes are ambiguous and multi-dimensional.

There is no clean event to meter.

2. Outcomes are not attributable to a single execution

AI systems are not linear.

They involve:

multiple steps
retries
tool use
retrieval
background processes

Which step produced the outcome?

You can’t isolate it.
And if you can’t isolate it, you can’t price it.

3. Outcomes are not observable in real time

Pricing systems require immediate measurement.

But outcomes often show up:

later
indirectly
or not at all

If the outcome can’t be observed at the moment of execution,
it can’t be used to control cost.

4. Outcomes cannot bound cost

This is the fatal flaw.

Cost is incurred before the outcome is known.

So pricing on outcome means:

compute is spent first
outcome is evaluated later
billing happens after the fact

That eliminates the ability to:

estimate before execution
approve before spend
enforce limits in real time

You’ve turned cost control into post-mortem accounting.

5. Cost per outcome inherits the same blindness as tokens

Even if you could define an outcome, you still haven’t solved the core problem:

you don’t know how much work it took to produce it.

So two identical “outcomes” can require:

vastly different compute
vastly different cost

Which means:

margins become unpredictable
providers subsidize unknowingly
constraints get introduced later to compensate

The system starts to behave exactly like token-based pricing:

subsidize → tighten → rebalance → repeat

The core mistake

Cost per outcome feels like progress.
It’s actually avoidance.

Instead of fixing measurement, it abandons measurement.

But you can’t control what you don’t measure.

The alternative: FLOP-Based Metering (FBM)

FLOP-Based Metering (FBM) measures AI usage based on the compute work performed, not the text produced.

It answers a different question:

Tokens: how much text was generated?
FBM: how much work was done?

Why FBM exists

Token-based systems introduce four structural problems:

Mismatch: identical token counts can represent vastly different compute workloads
Invisibility: agentic and background processes consume compute without generating tokens
Lack of control: tokens are measured after execution, so cost cannot be bounded in advance
Semantic blindness: all tokens are treated as equal, despite differences in meaning and computational effort

FBM replaces the proxy with the thing itself:

compute.

What “FLOP-based” means

FBM measures the computational work required to perform inference using floating point operations (FLOPs).

Instead of counting words, it measures:

the actual operations required to produce the result

This aligns measurement with resource consumption, not output.

FBM’s defining characteristic: compute alignment

In token systems:

measurement is indirect
cost is inferred from output
variability is hidden

In FBM:

measurement reflects execution
cost aligns with work performed
variability is visible

This alignment is what makes control possible.

How it works

For each request, the system evaluates the projected computational workload.

From this, it derives a compute-based measure that can be used:

before execution (to estimate and approve)
during execution (to monitor)
after execution (to reconcile)

The specific calculation methods are not disclosed, but the principle is clear: measurement reflects work, not words.

What measurement means in FBM

Measurement in FBM represents the economic cost of execution.

It is not:

a proxy for text length
a billing artifact
a post-hoc approximation

It is a direct representation of the compute required.

Why this matters

Without a compute-aligned unit:

cost cannot be predicted
cost cannot be approved
cost cannot be controlled

FBM enables:

pre-execution estimation
per-request visibility
enforcement of cost boundaries

In other words: cost discipline.

The bottom line

Cost per outcome tries to price the result.

FBM measures the work that produces the result.

One is an abstraction layered on top of a broken system.
The other fixes the foundation.

Final line

You can price the outcome.

But if you can’t measure the work,
you will miss the target.

FLOP-Based Metering, in detail.

– Published on Thursday, March 26, 2026