Based on OpenRouter’s Insights post. The figures below are OpenRouter’s own proprietary platform data, not independently verified; OpenRouter is a token marketplace with an interest in volume.
OpenRouter’s own platform data shows discounted GPT-5.6-family models multiplying token volume up to 13.8x — a strong elasticity signal, but the Jevons headline runs well ahead of what volume-and-share numbers can actually prove.
What Happened
In a post published this week on its Insights blog, OpenRouter — a token marketplace whose business grows with every API call routed through it — reported on a subsidized discount program OpenAI ran for models in its GPT-5.6 family between 27 July and 14 August 2026. All figures that follow are OpenRouter’s own proprietary platform data, self-reported by a company with a direct commercial interest in higher token volume; none of it is independently verified, and it should be read accordingly.
The setup is genuinely useful as a natural experiment. Two discounted models — which OpenRouter labels Terra and Luna in its post, names that are the company’s own program labels and should not be mapped onto specific product SKUs — saw list-price cuts of 20% and 80% respectively, then additional markdowns that brought their effective discounts to roughly 60% and 90% off. A third model, Sol, went un-discounted through most of the window, functioning as an implicit control. No per-token prices are published. What OpenRouter measured was token volume relative to each model’s pre-period daily average, and share of total tokens flowing through its platform.
The volume response was stark: Terra rose roughly 5.6x, Luna roughly 13.8x, and Sol roughly 1.1x. The discounted pair’s combined share of all OpenRouter tokens climbed from about 0.7% to about 7.8%. OpenRouter headlined this “Jevons paradox.” That framing is the part that needs careful handling — and it is where the analysis has to slow down before the structural read can begin.
The key insight: The experiment design is clean and the elasticity signal is real — deeper discounts produced sharply higher volume multiples while the un-discounted control barely moved. What the data cannot show is whether total AI spend rose, because it measures volume and share, not total consumption or total expenditure. Jevons paradox is precisely a claim about total spend rising despite lower unit prices. That claim is unproven here. What is proven is that AI-token demand is steeply price-elastic on a marketplace where switching costs are one API call.
The Structural Read
Strip away the Jevons framing and three structural dynamics emerge from the data, each worth taking seriously on its own terms.
Price elasticity is the commoditization flywheel. The relationship between discount depth and volume response is not linear — it is steep and convex. A 60%-off model did 5.6x. A 90%-off model did 13.8x. The control did 1.1x. That shape tells you the model layer is already behaving like a commodity: price is the dominant variable, quality differentiation is thin enough that developers will reroute without friction, and each incremental price cut unlocks disproportionately more usage. This is the demand side of the open-weight and price-war thesis made numerically visible. As the model layer races toward zero, volume does not grow proportionally — it multiplies. The question that follows is where the value goes when it does. The answer the data implies: to whoever sits on the flow. Routers, gateways, aggregators, and inference-layer intermediaries capture the volume that the model provider is effectively subsidizing. OpenRouter, which published this analysis, is itself one of those intermediaries — which is worth holding in mind when reading its interpretation.
Substitution versus net-new demand. The most important hedge in the data is also the easiest to miss. While discounted-model volume surged, Anthropic’s token volume on OpenRouter declined over the same window, and other OpenAI models surrendered roughly 1.9 percentage points of share. On a router, a price move is immediately visible to every developer making an API call, and re-routing is trivial. What looks like explosive growth for the winner is simultaneously a quiet decline for everyone else. The total pie — all tokens flowing through OpenRouter — may have moved only modestly. A large share of the 5.6x and 13.8x multipliers was not new demand conjured from nowhere; it was existing demand re-routed to wherever was cheapest that week. That distinction matters enormously for anyone reading this as evidence that cheap AI permanently expands the overall market. It may. But this data does not show it.
A subsidized promotion, not organic behavior. The durability test is the most damning gap in the Jevons reading. True Jevons dynamics are structural: lower unit cost → sustained expansion of use → permanently higher total consumption. What the post-program retention numbers show is the opposite of permanence. Roughly 32% of new users kept any usage after the discounts ended; approximately 18% maintained their in-program pace. The majority of the lift evaporated when the subsidy did. That is promotional elasticity — a well-understood and much less interesting phenomenon. Cheap prices pull enormous short-term volume, much of it borrowed from competitors, much of it transient. It is the marketing equivalent of a flash sale, not evidence of a structural shift in how much AI the world consumes.
The AI Value Chain — Business Engineer
When the model is a loss-leader, the money is in the flow
A world where a 90%-off model does 13.8x the tokens is a world where the model layer has become a loss-leader and the economic value has migrated upstream to the aggregation layer. OpenRouter sits on the flow and is both the observer and a beneficiary of that shift — which is precisely why its self-reported numbers should be read with that structural interest in plain view. The commoditization flywheel the data describes is real. The open-weight price-war dynamics driving it are real. The Jevons label is the costume, not the finding.
Three Implications
FOR ROUTERS AND AGGREGATORS — VOLUME IS THE BUSINESS
If deep discounts multiply token volume — even mostly through substitution — the intermediary routing that volume wins regardless of which model gets the traffic. OpenRouter, Vercel AI Gateway, and similar aggregation-layer players accumulate strategic leverage as the model layer races to zero. The open-weight and price-war barbell that is reshaping the inference market accelerates this dynamic: more price competition means more traffic to whoever makes routing frictionless. That is a durable structural position, independent of whether any single promotional surge is Jevons or not.
FOR MODEL PROVIDERS — PROMOTIONS BUY VOLUME, NOT LOYALTY
The retention numbers are the clearest signal in the dataset: ~32% kept any usage, ~18% held pace. Subsidized price cuts on a marketplace where switching is one API call produce traffic spikes, not customer relationships. For OpenAI or any provider running similar programs, the honest ROI question is whether the ~18% who stuck constituted enough durable volume to justify the discount cost — and whether that cohort is identifiable and buildable. Promotional elasticity is a real lever; confusing it with structural demand expansion is a planning error.
FOR ANYONE READING AI MARKET DATA — SOURCE STRUCTURE MATTERS
This is OpenRouter’s data, published by OpenRouter, about an event that made OpenRouter look like the indispensable infrastructure layer for AI token markets. That does not make the numbers wrong — but it is the mandatory prior before interpretation. Self-reported metrics from a marketplace whose revenue scales with volume should be weighted accordingly: useful as directional signal, insufficient as market proof. The Jevons framing is the clearest example of that gap: it is a bigger claim than the data supports, and it happens to be a flattering one for the publisher.
