The Price of Intelligence: A Quality-Adjusted Price Index for AI Services

URL SCAN: The Price of Intelligence: A Quality-Adjusted Price Index for AI Services
FIRST LINE: Economics > General Economics

The Dissection

This paper is an autopsy of AI price statistics. Matched-model accounting sees only a 0.10 log-point annual decline, while quality-adjusted accounting sees 0.73—87% of the decline is erased by treating improving models as unchanged goods. It then punctures the triumphalism: per completed task, buyer cost has stopped falling because reasoning models consume tokens faster than token prices decline. Nearly identical model rankings—0.998—can still produce a 0.49 log-point annual index shift after contaminated benchmarks are removed.

The paper is therefore measuring two diverging markets: seller-priced inference units and buyer-priced completed cognition.

The Core Fallacy

The paper’s limiting error, under the Discontinuity Thesis, is treating better price measurement as the central economic question. It is not. The decisive variables are substitution, control, deployment, and whether human labor remains necessary.

A flat completed-task price does not preserve the wage-to-consumption circuit. Firms can still acquire more capable cognition for roughly the same task budget and reduce headcount or eliminate human bottlenecks. Conversely, a falling quality-adjusted price alone does not prove mass displacement; the supplied abstract contains no direct evidence on labor substitution or ownership. The index sharpens the instrument, but it does not reach the organ DT says is dying.

Hidden Assumptions

• Benchmark response patterns are valid proxies for economically useful intelligence.
• A latent quality score can compare models across tasks, providers, and time.
• “Completed task” remains stable when reasoning depth and token consumption change.
• Public posted prices represent the relevant market, including discounts, bundles, self-hosted systems, and capacity constraints.
• Price and quality statistics can be interpreted without measuring ownership and control of AI capital.
• The post-2024 trajectory is durable rather than temporary pricing and model churn.
• Removing flagged benchmarks adequately controls contamination, despite leaderboard stability failing to guarantee index stability.

Social Function

Primary classification: partial truth. Secondary classifications: prestige signaling and transition management.

This is not pure copium. It exposes a genuine statistical blind spot and a real split between seller price and buyer task cost. But its technocratic frame can become ideological anesthesia: institutions can repair the index, announce higher measured productivity, and avoid the harder fact that productivity gains may accrue to sovereign owners while the majority lose productive necessity.

The Verdict

The paper finds the hidden pulse of the machine: intelligence is becoming cheaper per unit of capability even when a finished task is not. That may delay the fantasy of instantly free cognition, but it does not rescue post-WWII capitalism. It identifies a deployment friction, not a reversal of P1–P3. If quality keeps rising, ownership remains concentrated, and substitution follows, seller-side price declines become a transfer toward AI capital owners while task-level cost volatility serves as hospice care for the old labor circuit.