The Collapse in the Price of Inference: Why Intelligence Getting Cheap May Matter More Than Intelligence Getting Good

01 The Event and the Analytical Inversion

The observed facts are straightforward. New frontier models from Anthropic and OpenAI are being released in rapid succession, and each generation pairs greater measured capability with lower posted inference prices — the per-token rates that application developers pay to run these systems. The industry press describes this as a price war. That framing is not wrong, but it is analytically shallow. A price war is a competitive strategy; what is happening is better understood as a structural event in the economics of a new input.

The seed question posed for this analysis inverts the conventional narrative. Ordinary coverage asks: how smart are the new models? The economically interesting question is: what happens when a productive input — machine-generated analysis, drafting, and decision support — follows a steep cost-decline curve while its quality simultaneously improves? Historically, technologies that combine falling marginal cost with rising quality — electric power, computing itself, broadband — do not merely displace incumbent suppliers. They change the texture of downstream economic activity, because at some price threshold, uses that were previously irrational become rational, and whole application categories become newly viable.

This analysis therefore treats the frontier-model price competition as a case study in input-cost economics. It distinguishes observed facts (posted prices, released models) from reported claims (demand responses attributed to providers) from causal inference (what falling prices should do to market structure) from model-based projection and scenario. The central argument can be stated at the outset: the economically dominant variable is the price of inference, not the intelligence of models, because price determines the breadth of adoption, and breadth of adoption — not peak capability — is what propagates effects through the economy.

02 The Cost-of-Intelligence Framework: Price per Unit of Analysis

Artificial intelligence, as the reference record defines it, is the capability of computational systems to perform tasks typically associated with human intelligence — learning, reasoning, problem-solving, perception, and decision-making — developed as a field of research across engineering, mathematics, and computer science (source: Wikipedia summary — Artificial intelligence). For three decades after that field's founding, the unit cost of machine intelligence was dominated by research expenditure: enormous fixed costs to build systems, negligible deployment. The current phase is different. Foundation models have converted intelligence into a metered commodity. A developer does not buy an AI system; she buys tokens — units of text processed — at posted prices, the way a factory buys kilowatt-hours.

That metering is what makes economic analysis possible. When a productive input is metered, three classical questions become answerable: What is the price trend? What is the elasticity of demand with respect to price? And who captures the surplus when price falls? The posted-price record across the frontier providers shows repeated, large, headline-grabbing cuts in per-token rates coinciding with new model releases — with newer generations frequently priced well below older ones of comparable or greater capability. These are observed prices, not projections: they are the public list rates at which access is sold.

The mechanism behind the falling price is worth decomposing, because it determines whether the decline is durable. Cost per token is approximately a product of four factors: the capital cost of the compute cluster (depreciated over served tokens), the energy and cooling required per unit of computation, the algorithmic efficiency of the model at a given capability level, and utilization — how fully the expensive capacity is kept busy. Declines in posted prices imply declines in at least one of these components. Algorithmic efficiency gains — techniques that achieve a given capability with fewer floating-point operations — are the most cited driver, and they compound: each efficiency gain lowers the cost floor for every subsequent model. But capital costs are simultaneously rising as providers build larger training clusters, which means the price cuts are being financed by spreading enormous fixed costs across a volume of inference that must grow steeply for the arithmetic to work. This is the classic infrastructure-business tension, and it is the load-bearing assumption beneath every posted price cut.

03 The Jevons Question: Does Cheap Inference Mean Less Spending?

The most common wrong inference about falling token prices is that they imply falling AI spending. The nineteenth-century economist William Stanley Jevons observed that improvements in the efficiency of coal use did not reduce coal consumption — they increased it, because cheaper effective use expanded the set of profitable applications. The modern formulation: when the effective price of a productive input falls, total expenditure on that input rises if demand is sufficiently elastic. The relevant analytical question is therefore not whether per-token prices fall — that is observed — but whether demand for tokens is elastic enough over the relevant horizon for the Jevons dynamic to dominate.

Here the evidence categories matter. Reported claims from providers point toward high and rising inference volumes as prices fall — consistent with elasticity greater than one. But three structural features of this market counsel caution before extrapolating. First, the elasticity is use-case specific: token demand from a code-completion assistant responds to price; token demand from a researcher running a one-time analysis barely does. The aggregate elasticity depends on the mix, and the mix is shifting toward volume-heavy agentic use in which each user session generates many internal model calls. Second, part of the observed volume growth is a composition effect: as models become more capable, each unit of work requires more tokens of intermediate reasoning — a chain-of-thought process literally spends tokens on thinking — so revenue per unit of task output can fall even as token volumes soar. Third, elasticity is not a constant; it is a curve, and the elasticity measured at today's prices may not describe demand at prices an order of magnitude lower.

The honest statement of the inference: falling posted prices plus rising provider revenues (where reported) are consistent with a Jevons regime — but the aggregate elasticity of demand for machine intelligence has never been measured at scale, because the commodity is barely three years old as a metered input. This is a genuine unresolved empirical question, and it determines the second-order economics: if the Jevons dynamic dominates, total spending on inference rises and the constraint shifts to compute supply; if it does not, falling prices translate directly into provider-margin compression and an eventual shakeout.

04 Who Captures the Surplus: A Value-Chain Accounting

When the price of an input collapses, the savings do not evaporate — they are distributed. The economics of surplus capture follow a pattern familiar from earlier input-cost transitions: whoever controls the scarcest complementary asset captures the rent, and the commoditized layer captures the least. If inference becomes the commoditized layer, the surplus from cheaper intelligence flows to the complements. Four candidate captors can be identified, in descending order of how directly the observed price cuts hand them value.

First, application developers: falling per-token costs lower the marginal cost of running a product that embeds model calls. A service whose unit economics were negative at old prices becomes viable at new ones, so the price cuts act as a subsidy to the entire application layer — the same developers whose token literacy the source video for this analysis describes from a practitioner's standpoint (source: source video — Most devs don't understand how LLM tokens work, Matt Pocock). Second, end consumers of software, who capture surplus only insofar as competition among application providers forces the input savings into lower prices or better free tiers; where application markets are winner-take-all, developers retain more of the surplus. Third, the model providers themselves, who are voluntarily surrendering margin on inference — an apparent anomaly in surplus accounting. The rationalizations are strategic: posted-price cuts can be a demand-elasticity play (volume recovers margin), a competitive-moat play (pricing below cost to discipline rivals), or an ecosystem play (cheap tokens make a provider's tooling standard). Fourth, the suppliers of the underlying physical inputs — accelerator chips, power, and data-center capacity — who capture value only if the Jevons dynamic holds; if aggregate token demand grows faster than price falls, the true bottleneck migrates from models to electrons, and the scarce-rent shifts accordingly.

This accounting explains a puzzle in ordinary coverage: if inference is a price-war business with collapsing unit prices, why is capital flooding into it? The answer is that investors are not necessarily buying inference margins. They are buying positions in the layer where scarcity will concentrate once the model layer commoditizes — distribution, proprietary data, and the physical supply chain. The falling price of intelligence is, in accounting terms, a transfer from the model layer to whoever owns the complements.

05 Second- and Third-Order Effects: The Migration of the Bottleneck

The first-order effect of cheaper inference is straightforward: more applications, more tokens, more embedded machine reasoning in ordinary software. The second-order effects are where the system gets interesting, and they follow from the migration of the binding constraint. When the cost of intelligence falls, the constraint on any given workflow shifts from the reasoning step to its complements: data access, integration, verification, and liability. In economic terms, the complementary factors become relatively more expensive, and their owners gain bargaining power.

Trace one causal chain. Cheap inference → automated agents that draft, code, analyze, and correspond at near-zero marginal cost → a rise in the volume of machine-generated content and communication → a decline in the credibility value of unprompted text and a rise in the value of verification mechanisms → third-order: institutional adaptation as courts, universities, and firms redesign their evidence standards around the assumption that competent text is free. Each arrow is a causal inference, and the later links are speculative — but the mechanism is the same one that followed desktop publishing (the cost of typesetting collapsed; typography stopped signaling professionalism) and stock photography (the cost of a plausible image collapsed; generic imagery stopped signaling effort). What the frontier-model price war does is collapse the cost of plausible analysis and articulate prose. The third-order institutional consequence — what stops functioning as a signal of competence when writing well is free — is likely to be a larger economic story than any single model release.

A second chain runs through the labor market. Cheap machine reasoning is not a uniform shock: it reduces the cost of tasks that can be specified as text-in, text-out (drafting, summarizing, translating, coding boilerplate) while leaving physical, relational, and judgment-under-liability tasks relatively untouched. The wage effects therefore concentrate in occupations whose value proposition is largely cognitive-textual production — entry-level knowledge work above all, because the tasks juniors perform are the most token-like. This is a model-based projection, not an observed fact; but it inverts the usual skill-bias story of recent decades, in which technology raised the relative wage of the educated. A technology that cheapens the output of educated labor is a different kind of shock, and its distributional consequences — potentially compressive at the top of the wage distribution rather than amplifying — are among the most important open questions in labor economics.

06 Historical Precedent: Lighting Up the Grid, and the Difference That Matters

The closest historical analogue to a metered intelligence input is not software; it is electric power. Between the 1890s and the 1920s, the price of a kilowatt-hour fell dramatically while the capability of electric appliances rose. The economic historian Paul David's celebrated analysis of the productivity slowdown showed that electricity's effects arrived not when the price fell but when factories — built around the steam layout of a central shaft — were physically reorganized around the new input's properties. The lesson: input-cost revolutions require complementary reorganization before they show up in productivity statistics, and the reorganization is measured in decades, not quarters.

What is similar between electrification and the current moment: a metered input, falling real prices, capability improving alongside, early adoption that substitutes the new input into old workflows (the electric light replacing the gas lamp; the chatbot answering the email), and a measured productivity effect that lags the price effect. What is different, and why the difference matters: first, electricity's cost floor was physical and its price decline eventually flattened as generation hit thermodynamic and fuel constraints; token prices are falling on an algorithmic- plus silicon- cost floor whose bottom is not known — it is conceivable that near-frontier reasoning becomes nearly free, which has no electrotechnical precedent. Second, electricity could not copy itself; knowledge goods embedded in models improve for every user simultaneously when the frontier advances, so quality gains compound across the whole installed base in a way that kilowatt-hours never did. Third, the institutional complement — the factory reorganization — was slow but uncontested; the equivalent reorganization for machine reasoning runs into liability rules, professional licensing, and verification norms that may actively resist it. The electrification precedent therefore cuts both ways: it predicts a lagged, reorganization-gated productivity payoff — and it warns against assuming the lag will be tolerated by markets that have priced in the payoff arriving faster.

07 Scenarios and Indicators to Watch

N43 offers three scenarios for the cost-of-intelligence market over the next several quarters, each with observable triggers. These are scenarios, not forecasts, and no probabilities are asserted beyond reasoned weights.

Scenario A — Stabilization of the price curve. Posted prices stop falling as providers reach a cost floor dominated by capital and energy rather than algorithms; competition shifts from price to reliability, latency, and enterprise integration. Trigger: two consecutive model cycles with roughly flat list pricing. Indicator: the price differential between top-tier and mid-tier models narrowing instead of widening, and provider communications shifting from price leadership to quality-of-service claims.

Scenario B — Persistence of the decline. The current pattern continues: each generation ships cheaper per token than the last, volume grows steeply, and the Jevons regime holds. Trigger: new releases whose posted prices undercut predecessors by large factors. Indicators: aggregate industry inference revenue rising despite per-token declines; utilization pressure visible as rate limits, queuing, or capacity announcements; accelerated data-center and power contracting. The second-order consequence to watch in this scenario is the migration of the bottleneck to electricity and grid interconnection — the complements capture the rent.

Scenario C — Structural break: free-tier intelligence. Perceptibly free, high-quality inference becomes the default consumer tier — the way search is free — with monetization migrating to advertising, enterprise contracts, or ecosystem lock-in. Trigger: a major provider making a frontier-grade model free at scale, sustainably. Indicators: the appearance of inference-heavy consumer products priced at zero with no visible metering; consolidation among mid-tier model providers who cannot subsidize free tiers; regulatory attention to concentration in whatever complement becomes monetizable. This is the scenario in which the surplus fully exits the model layer.

Indicators to watch across all scenarios: posted per-token list prices at each model generation (the cleanest observable); the ratio of revenue growth to token-volume growth at major providers where disclosed (a direct test of the Jevons question); data-center and power-procurement announcements (the physical constraint); the agentic application pipeline — how many products ship whose unit costs depend on cheap inference; enterprise pricing structures shifting from metered tokens to flat-rate contracts (a sign that providers are internalizing the price war); and any provider reporting inference margins explicitly. Each indicator maps to a specific mechanism identified above, which is what makes it worth watching.

08 The Bottom Line

What we know: New frontier models from Anthropic and OpenAI pair greater capability with lower posted per-token prices; per-token inference has become a metered commodity with a steeply declining list-price trend. Falling input prices with rising quality historically transform downstream markets rather than merely reallocating supplier share.

What we think we know: The demand side is responding elastically — inference volumes are growing fast enough that falling unit prices have not collapsed total spending, which points to a Jevons regime. The surplus from cheap intelligence is being transferred from the model layer toward complementary assets: application ecosystems, proprietary data, and the physical supply chain of compute and power. The binding constraint on the industry is migrating from model capability toward energy and capital.

What we do not know: The aggregate elasticity of demand for machine intelligence — never measured at scale for a commodity this young — and therefore whether the Jevons regime is durable. The location of the true cost floor: whether near-frontier inference can approach zero price, which no historical input transition precedent answers. And the third-order institutional question of what happens to the signaling value of articulate human work when its production cost collapses.

Signal versus noise: Individual model releases are noise; the posted-price time series is signal. A structural read of this moment should track the price curve, not the benchmark leaderboards — because the price curve is what determines who can afford to build, and who builds determines what the economy becomes. The frontier-model price war, on this reading, is not a marketing contest between two laboratories. It is the visible surface of a repricing of reasoning itself, and repricings of that depth historically take a decade to metabolize. The most important numbers in AI right now are not benchmark scores. They are the list prices, and which direction they are going.