NEW IN THIS EDITION
Every result in this document divides by m, so m needs a unit. "Units of completed work" cannot tell a one-line
fix from a framework migration, and the frontier case is overwhelming for one and absent for the other. The unit
used throughout is the accepted work unit, and the variable that decides the premium is not the
size of the artifact but how many autonomous steps run between two acceptance gates.
Definition · the accepted work unit
One change that (a) passes a stated acceptance gate — a test suite, a review, a
regression run, a sign-off; (b) survives a set window without being reverted — 30 days
here, so rework is not counted twice; and (c) is sized in expert-hours — the time a
skilled human would take to produce the same accepted change. Condition (b) is what turns generated output into
accepted output (limit 4 in driver 2). Condition (c) gives every unit the same currency, so a migration counts
for more than a typo fix and the count cannot be gamed by splitting work into small pieces.
Completion probability by unit of output
P(unit passes its gate with no re-drive) = pN · N = autonomous steps between gates (illustrative per class) · label = frontier advantage, 99% over 97% per step
| Unit class | Acceptance gate | N |
97% | 99% | Edge |
|---|
N is the interval between gates, not the size of the artifact
A full application is not one unit. Nobody runs a thousand steps unattended: the work is cut into gated pieces
— spec, features, integration, acceptance testing — and each gate resets N. The
right-hand bar in the chart is therefore a limit case, not a deliverable: at 200 ungated steps both models fail
almost every time, which is exactly why no one buys output that way. What a buyer actually procures is a stream
of gated units, and the premium is priced unit by unit. This also makes Note 2's caveat precise: human checkpoints
do truncate the chain, and the place they truncate it is the unit boundary.
Tier 1 · Task
Activity
Classify, extract, route, a one-line fix. Short N, so pN barely separates
models, and value is roughly linear in volume. This is the work the open-weights and mid-tier
lanes clear.
Lane: 1–2 · price per token matters
Tier 2 · Deliverable
Artifact
A feature, a merged change, an upgrade or migration. Pass/fail at a gate, long N.
Drivers 1 and 4 both bite here: compounding sets the completion rate and the threshold sets
whether partial work is worth anything.
Lane: 3 · the premium zone
Tier 3 · Outcome
Result
Process improvement: cycle time, defect rate, cost to serve. Not a chain of steps but
the Y that deliverables, the X's, move. The frontier acts here only indirectly, by raising how
many deliverables can be pushed at the outcome.
Mode: B · bounded by opportunity set
Where the premium pays off — long gated intervals
The premium pays off where a unit is long between gates and the gate is binary. Upgrades and
migrations are the cleanest case: framework end-of-life, runtime and dependency upgrades, security
patching, legacy modernization. A module either builds and passes its full regression suite or it does not, the
gate is checked by machine rather than by a reviewer's judgment, the interval is long, and most enterprises hold
a large backlog of exactly this work. At N ≈ 100 the two-point reliability edge is a
7.7× difference in units that clear the gate unaided. Features follow at 2.8×.
Bounded tasks sit just above the 4.0% bar — about 6% at N ≈ 3 — close enough that a
cost-tuned mid tier or ordinary commercial terms erase it.
What the definition makes measurable
With units sized in expert-hours, m = accepted expert-hours delivered ÷ labor-hours paid.
That puts m on the same labor denominator as f, so (1 + f) / m
becomes something a firm can measure from its own merge and revert history rather than an illustration. The
identity, the break-even E = (fF − fO) / (1 + fO),
and the 4.0% bar are all unchanged; what changes is that m can now be read per unit class
instead of as one blended average, which is the averaging trap limit 2 warns about. Sizing tasks by the time a
skilled human needs is the same yardstick that task-length ("time horizon") evaluations of AI agents use.
The single largest driver, and the one invisible on single-turn benchmarks. Success on an N-step agentic task
approximates pN. A 2-point per-step reliability edge is nearly invisible in
a benchmark table and decisive by step 50.
Task completion probability vs agentic horizon
Y = P(all steps succeed), % · X = steps in the task (log scale)
| Per-step reliability | 10 steps | 25 steps |
50 steps | 100 steps |
|---|
Why this dominates
Moving per-step reliability from 97% to 99% — a two-point delta a composite index would
round away — is a 2.8× swing in completion at N=50 and 7.7× at
N=100. That is the boundary between "the agent ships the PR" and "an engineer babysits and re-drives."
Read against driver 2: the completion-ratio gain from that edge is 2.1% at N=1, 4.2% at N=2, 22.6% at N=10 — it clears the 4.0% break-even bar from N=2 onward, by a widening margin. The bridge from completion probability to m is accepted output (defined in section 0; limit 4 in driver 2): a task that fails to complete ships no feature and adds nothing to m.
REVISEDREFRAMED
Cost per unit of output is (1 + f) / m, where f is model spend as a
fraction of loaded labor and m is the throughput multiplier. f is additive; m is a divisor. A
small numerator penalty buys a large denominator gain, which is why the frontier wins while spending several times
the tokens.
Reading the formula
Worked example · one engineer, one year
Illustrative accepted work units (section 0) — not a measurement
Why the asymmetry is the whole argument
Elasticity of unit cost with respect to each input
| Input | Elasticity | At f = 5% | Reading |
|---|
Scenario comparison
| Scenario | f · spend | m · throughput |
Unit cost | vs baseline |
|---|
Throughput edge the frontier must deliver to break even
Y = required advantage over open weights, % · X = frontier model spend as % of loaded labor · each curve is a different open-weights spend baseline · shaded band = revised realistic spend
Break-even in closed form
What the identity cannot see
Revised spend band — the earlier 1–2% was anchored too low
A saturated agentic engineer running parallel sessions lands nearer
$500–$1,200/seat/month, or
2.4–5.8% of $250k loaded labor, with a fat right tail
(MED — inferred from agentic usage patterns, not a surveyed figure). Raising f does not weaken the
case: at 5% frontier spend against 1% open weights, break-even is a 4.0% throughput
edge. Even at 15% spend it is only 13.9%. The token-price argument was never
load-bearing.
Jevons drift — and where the frame breaks
As inference gets cheaper per unit of work, consumption rises faster than unit cost falls, so total spend
goes up — Gartner predicted in June 2026 that AI coding costs will exceed the average developer
salary by 2028. The 2.4–5.8% figure will drift upward, and that drift
is a health signal rather than a cost problem. Past some point "tooling as a percentage of labor" stops being
the right accounting frame entirely: compute becomes a factor of production alongside labor
rather than an overhead line charged against it. That is when the procurement conversation changes shape.
NEW IN THIS EDITION
A throughput multiplier can be banked two ways: hold output and cut headcount, or hold headcount and expand
output. They are identical on unit cost and completely different on total value. Illustration
below: a 100-engineer team at m = 1.45, f = 5%.
Two ways to bank the same multiplier
Indexed to baseline = 100 · cost, output, and cost per unit of output
Mode A
Reduce headcount
Value is capped at the labor line — you recover at most the
salaries eliminated, and the gain terminates. Note that f itself does not move: N/m engineers at the same per-seat spend scale
numerator and denominator together. What shrinks is the absolute AI budget, which can make a
high-leverage line item look small and invite under-investment — a perception risk, not arithmetic.
Forced when: demand is the binding constraint
Mode B
Hold headcount, expand output
Value is bounded by opportunity set, not by cost. Unbounded upside where
a project backlog exists — additional revenue-generating work rather than a one-time cost recovery.
Available when: engineering capacity is the constraint
The choice is a diagnostic, not a preference
Both modes land at 72.4 on cost per unit of output — the
arithmetic cannot distinguish them. What separates them is which constraint actually binds. Most software
organizations claim engineering capacity is their limit, often accurately; firms that reflexively take Mode A
are usually revealing that their real constraint was never headcount. That is a strategy
disclosure, not a procurement decision.
"Open weights reach ~79% of proprietary" (87% in June) implies linear value transfer. On threshold-gated work — the task
either clears the bar or it doesn't — delivered value is sigmoid, and the index points
between open weights and the frontier straddle the knee.
Delivered value vs model capability
Illustrative: assumed linear transfer vs threshold-gated response · knee at ~88 · marks at the June (87%) and October (79%) gaps
The composite-average trap, priced
On this curve June's 87% of frontier capability delivered ~38.9% of the value, not 87%.
October's wider gap, 79%, sits below the knee entirely at ~1.7%: the exact figure is an
artifact of the chosen curve, but the direction is not — a wider gap on gated work forfeits value far
faster than the index suggests. The benchmark isn't wrong — a mean over ten evals is simply the wrong functional form for work gated on a pass
threshold. The easy 80% of tasks were automatable two model generations ago; the residue is exactly what sits on
the knee. Illustrative model — LOW confidence on curvature, HIGH on direction.
Open weights are free; serving them is not. Fixed cost is amortized over tokens actually
produced, so effective unit cost is governed by utilization — and enterprise diurnal load runs
5–10× peak-to-trough.
Effective cost per 1M output tokens vs fleet utilization
Modeled: 4× 8×H200 nodes + 3 ops FTE · shaded band = realistic enterprise utilization
Model assumptions — load-bearing, tagged
Infra $1.26M/yr (4 nodes × $36/hr, 8×H200 rented at $4.50/GPU-hr, the low
end of October 2026 on-demand pricing — MED) plus ops
$900k/yr (3 FTE × $300k loaded — MED) = $2.16M/yr fixed.
Peak capacity 5,000 output tok/s per node for a ~100B-class MoE under heavy batching — LOW
confidence; this is the swing variable. Halve throughput and every crossover doubles. Reference
prices are October 2026 list output prices. The qualitative result: self-hosting beats a flagship
price almost immediately, a mid-tier price only at the top of the realistic band, and a cheap
open-weights API not at all — open-weights API prices have fallen below what a private
fleet can match, so the case for self-hosting is now privacy, residency, or control, not unit cost.
Eight drivers, ordered by how much of the premium they explain, with basis and confidence on each.
Lane 1 · Open weights
Bounded, high-volume, private
Classification, extraction, routing, embedding. Error-tolerant, short-horizon, and
cheap enough per token that capability is not the constraint. The open-weights case holds cleanly
here — usually through a hosted API, with self-hosting reserved for privacy and residency.
Also: narrow fine-tunes that beat general frontier on one task
Lane 2 · Mid-tier proprietary
General reasoning, good-enough
Where the dominated-flagship trap lives. A cost-tuned proprietary
tier now scores above every open-weights model at a comparable cost per task (Oct 2026: 48 at $0.21
against the best open model's 46 at $0.13) — and at a fraction of a flagship's.
Trap: max-effort by default on work a mid tier clears
Lane 3 · Frontier
Long-horizon, calibration-critical
Agentic chains, must-not-hallucinate, threshold-gated. pN and the sigmoid knee
are both load-bearing. Token cost is the small line item.
Plus: indemnity, compliance, harness — often dispositive
Net
Enterprises don't pay the premium because they're bad at arithmetic — they pay it because the arithmetic
they're doing has a different denominator, and because f enters additively while m divides. Where open weights
genuinely win is a real and sizeable quadrant, but it's bounded, high-volume work, not
general reasoning. The overpayment that actually exists is failure to segment traffic, and it
shows up inside the proprietary cohort as often as across the closed/open line.
Everything in this document that is a choice rather than a measurement, stated plainly,
plus what changed in this edition and why.