The "50% off, 14x the tokens" number is real, but our call is that it measures users switching onto OpenRouter, not new demand. OpenRouter's own earlier study of its platform found price sensitivity about one-twentieth of what 13.8x implies.

Key takeaways

In late August a number went around the industry. OpenRouter is a routing platform: developers use it to compare prices and send requests across several hundred models from dozens of providers. OpenRouter posted that after a heavy discount, token usage of OpenAI's GPT-5.6 Luna model on OpenRouter grew 13.8x against the weeks before the discount, and OpenRouter labeled the jump a case of the Jevons paradox. Luna is one of three models in OpenAI's GPT-5.6 family, alongside Sol and Terra. One well-known investor reposted it, remarking that a 50% price cut for 14 times the volume was a bit extreme. The number was then put to work on a much bigger claim: price cuts will hold compute demand up, so data centers have nothing to worry about.

Our call: the number is not fabricated, but it measures which route users bought through, not how much more the world bought in total. And the design of this discount made the two impossible to tell apart, because the discount existed only on OpenRouter. For the whole discount window, the same Luna cost exactly half as much bought through OpenRouter as bought anywhere else. A route that costs half the price can grow fourteenfold without anyone in the world using a single additional token.

The essay runs in five steps:

① Take the discount apart into its two layers. One was market-wide, the other was a single channel's; that structure is why the numbers cannot separate "moved over" from "new demand."

② Put OpenRouter's two elasticity figures from two of OpenRouter's own studies side by side, and trace where the gap of twenty times or more comes from.

③ Redo the share breakdown OpenRouter offered in response to the switching objection, and find the condition under which it holds.

④ Lay out a counterintuitive structure: the smaller and less representative the channel, the more easily "moved over" explains the whole multiple.

⑤ Name the one reading in this episode that can carry a Jevons argument. That reading is the 18% of users who kept up their discount-period pace after the channel discount ended, not the 13.8x multiple.

What does this add to yesterday's daily brief on the same item? Three things. First, this essay rechecked yesterday's call, and half of it did not hold. Yesterday's brief said nobody had published how much of the jump came from users switching routes. That sentence was wrong: OpenRouter did publish a share breakdown this time, but the breakdown is miscalculated; and SemiAnalysis, a research firm that tracks the AI compute industry, had independently described the same structural problem ten days before OpenRouter's post. The core of yesterday's call survives; the clause "nobody has published" is overturned. Second, the price change behind 13.8x turns out to be two separate cuts, not one. The discount was not one cut but two stacked together: on July 30 OpenAI cut Luna's public list price by 80%, market-wide and permanently, while OpenRouter ran a separate 50% discount on OpenRouter alone, layered on top of the list price. Taken apart, the first is a price cut everyone received; the second is purely a channel price gap. Third, this essay fixes one condition that would prove it wrong, and the data for it already exists. From July 27 to 29, only the channel discount was in force and OpenAI's list price had not yet moved. The usage change over those three days is therefore a clean controlled comparison. OpenRouter holds that daily data, and publishing it would settle this essay one way or the other in a single step.

The "per-token economics flip" line in the tree above is a fight over whether model companies get to keep their new margin on the tokens they sell. This essay works on a more basic question underneath that fight: whether the yardstick everyone uses to judge "will a price cut buy enough volume" measures what it claims to measure.

First: the discount had two layers, and only one was market-wide

The timeline comes first, because every later step stands on it.

• July 27: OpenRouter announces 50% off GPT-5.6 Terra and Luna. The post reads, verbatim: "GPT-5.6 Terra & Luna are 50% off exclusively on OpenRouter!" (OpenRouter's official account).
• July 30: OpenAI cuts its own public list prices, Luna by 80% and Terra by 20%. Luna goes from US$1 per million input tokens and US$6 per million output tokens to US$0.20 / US$1.20; Terra goes from US$2.50 / US$15 to US$2.00 / US$12 (InfoWorld, CNBC). This is the standard API list price, not a limited-time promotion, and it applies to every customer through every channel.
• The same day, OpenRouter states that its 50% stacks on top of the new list price: "GPT-5.6 Terra and Luna just got price cuts from @OpenAI. OpenRouter is cheaper still. Our 50% exclusive discount applies on top, bringing Luna to 0.1/M in 0.6/M out and Terra to 1/M in and 6/M out." That is Luna at US$0.10 per million input tokens and US$0.60 output, and Terra at US$1 input and US$6 output (OpenRouter's official account).
• August 14: the channel discount ends. OpenAI's new list price stays.

Lining up the four rows produces the single most important piece of arithmetic in this essay:

For the entire discount window, the same model and the same tokens cost exactly twice as much bought anywhere other than OpenRouter.

That fact decides whether 13.8x can be used to talk about Jevons. The Jevons paradox says that when something gets cheaper, the world uses more of it in total. It is a claim about total quantity. But when one route becomes half the price of every other route, the first thing that happens is not that someone uses more. It is that people who were going to use it anyway change route. On OpenRouter's counter the two events look identical: both read as "more Luna tokens on the platform."

A detail that is easy to skip past: OpenRouter's users came to the platform precisely in order to compare prices and switch routes. The platform lays out the prices of more than four hundred models from more than sixty providers in one interface; switching is the feature it sells. So even granting that the typical developer finds changing providers a hassle, this group is the slice of the whole market with the lowest switching cost and the highest willingness to switch. Taking their response to a channel price gap as a proxy for the whole market's response to absolute prices biases the result in one fixed direction: it can only overstate, never understate.

Why this is not nitpicking

The objection runs: some of it could genuinely be new usage. Yes, and some of it surely is. The question is not whether, but how much. And this experiment's design makes the two parts inseparable.

OpenRouter knew it had to answer this challenge, and it offered a control that looks reasonable: GPT-5.6's third model, Sol, had no discount over the same period, and Sol's usage moved only 1.11x. That control does prove one thing: what happened to the discounted models was caused by the discount, not by the platform's organic growth. That point stands, and this essay does not contest it.

But the Sol control cannot prove the other thing. It cannot say how much of the extra Terra and Luna volume was new and how much moved over from another route. Sol had neither a channel price gap nor a list-price cut, so it is a "nothing happened" control, not an "only one of the two happened" control. Separating the two parts takes a window in which only the channel price fell and the list price did not. That window exists: July 27 to 29. Section five returns to it.

Second: one company, two studies, and elasticities twenty times apart

Now convert 13.8x into what economists call price elasticity: for each 1% cut in price, how many percent does usage rise.

After July 30, Luna's price on OpenRouter was one-tenth of its original price (an 80% list-price cut, then 50% off). Usage rose 13.8x. Dividing the logarithms gives an elasticity of about −1.14. Weighting by the actual days in the window (three days at 50% off only, then sixteen days at one-tenth of the original price), the average price paid was 0.163 of the original, for an elasticity of about −1.45. Terra's usage grew 5.6x, which by the same method is about −1.88 (end-of-window price) to −1.96 (day-weighted) (our calculation throughout).

Whichever figure is used, the order of magnitude is "price down 1%, usage up more than 1%": demand very sensitive to price.

Now the same company's other study. Early this year OpenRouter published a usage study spanning thirteen months and more than one hundred trillion tokens (State of AI, covering 2024-11-03 to 2025-11-30, also posted to arXiv). One sentence in that study:

"The trendline is nearly flat, indicating that demand is relatively price-inelastic; a 10% decrease in price corresponds to only about a 0.5–0.7% increase in usage."

As an elasticity: −0.05 to −0.07.

Same company, same platform: one study says −0.06, the other implies −1.14 to −1.96. A gap of twenty to thirty times.

This is not a contradiction and nobody is lying. Both numbers are right, because they measure two entirely different things:

• The −0.06 compares different models side by side at the same moment: how much usage differs between expensive models and cheap ones. It is closer to "how the whole market responds to price," but it carries a large confound. Expensive models are usually also stronger, so "expensive" and "more useful" are mixed together, which flattens the measured slope. The −0.06 is an underestimate.
• The −1.14 makes the same model cheaper on one route: users do not have to change any of their demand, just one line of configuration, to get the half-price rate. It mostly measures "how easy switching is," not "how big demand is." It is an overestimate.

The true market elasticity sits somewhere between the two, and this episode did not narrow that range. The figure the industry is currently making decisions on is the top edge of it.

The two numbers do not cover the same period (one runs from late 2024 to late 2025, the other covers July and August 2026), and the model mix is entirely different, so neither can directly overturn the other. But that is the point. Two numbers produced by the same company, from the same platform, differ by a factor of twenty on the design of the measurement alone. "This time the measurement read 13.8x" is then nowhere near enough to carry the inference "so price cuts will hold compute demand up."

Third: the share breakdown OpenRouter offered in response turns on a number OpenRouter did not print

OpenRouter evidently anticipated the objection that the volume came from somewhere else, and included a share breakdown in the post. The original reads:

"Terra/Luna went from 0.7% to 7.8% of all OpenRouter tokens between the pre-period and the program, a gain of 7.1 share points."

"Competitors gave up 5.3 points and other OpenAI models gave up 1.9 points over the same comparison, so roughly three quarters of the gain came from outside OpenAI."

(OpenRouter's official explanation)

OpenRouter concludes that roughly three quarters of the gain came from outside OpenAI rather than from OpenAI eating into its own volume.

That reading holds only if OpenAI's other models held about 27% of the platform before the discount. Public readings put OpenAI nearer 8%. The check below needs only pencil and paper.

Start with the "7.1 points." It is an identity and carries no information. A share is a ratio whose denominator is total platform volume. If Terra and Luna's usage explodes while every other model's absolute usage stays exactly the same, total platform volume grows, and every other model's share is diluted automatically. The dilution adds up to exactly the 7.1 points Terra and Luna gained.

The arithmetic: suppose total platform volume before the discount is 100 units, with Terra and Luna at 0.7 and everything else at 99.3. Hold every other model's absolute volume fixed (99.3 does not move), let Terra and Luna grow to X, and require that X's new share be 7.8%:

X ÷ (X + 99.3) = 0.078 → X ≈ 8.4

Terra and Luna grow from 0.7 to 8.4, roughly twelvefold, the same order of magnitude as the published multiple. The new total is 107.7. The other models' share goes from 99.3% to 99.3 ÷ 107.7 = 92.2%, down 7.1 points (our calculation).

Without a single token changing route, "the other models lost 7.1 points combined" still holds. The total proves nothing. It is division.

The informative part is the split: 5.3 against 1.9. Under zero switching, the dilution should fall on each group in proportion to its original share. So the question can be run backwards: for OpenAI's other models to absorb exactly 1.9 points, what share of the platform must they have held before the discount?

7.1 × (OpenAI's other models' share ÷ 99.3) = 1.9 → share ≈ 26.6%

OpenRouter's reading that "three quarters came from outside" holds only if OpenAI, excluding Terra and Luna, held about 27% of the platform before the discount.

What did OpenAI actually hold? Public readings point far lower. Several compilations put OpenAI at about 8.4% of OpenRouter's platform share in June 2026 (StockAlarm compilation); over the same period Chinese open-weight models took 40% to 60%, and the three big US labs (Google, OpenAI and Anthropic) together fell from about 70% a year earlier to about 30% (Dataconomy, TechBriefly).

Rerun the arithmetic with 8%:

The reading: of the 7.1 points, about 5.8 points are pure dilution produced by division, and only about 1.3 points represent anything actually changing hands. And that 1.3 points comes almost entirely from OpenAI's own other models. Competitors were not raided at all; they lost less than pure dilution would predict, which means their absolute usage very likely grew during the discount period (our calculation).

That is the opposite of the post's conclusion. The post says three quarters came from outside OpenAI; net of dilution, the identifiable switching happened almost entirely inside the OpenAI family.

The reversal rests entirely on one number, OpenAI's share of the platform before the discount, and the 8.4% used here is a secondhand compilation, not OpenRouter's official figure, on a platform whose shares move very fast (one dataset has OpenAI going from about 2% to about 8% within weeks). If that number is really 27%, the post's reading holds and this section is void. OpenRouter certainly has that number, and did not print it in the same post. That is this essay's point. The one number that decides which way the breakdown points was not published, and without it the breakdown has no direction.

Fourth: the smaller the channel, the more easily switching explains the whole multiple

If OpenRouter carries about 1% of Luna's global usage or less, switching alone can explain the whole 13.8x. The table below shows why.

The common intuition about a small sample is noisy but unbiased: a small sample just means a large error bar, and more measurements fix it. Here that intuition is wrong. The smaller the channel, the larger the bias, and its direction is fixed.

The reasoning: let s be OpenRouter's share of Luna's worldwide usage. For switching to explain 13.8x in full, OpenRouter must pull in, from outside the platform, 12.8 times its own original volume. What fraction of the outside pool is that? The formula is 12.8 × s ÷ (1 − s):

The smaller the channel, the larger the pool relative to it, and the less effort fourteenfold takes. Conversely, if OpenRouter carried a fifth of Luna's global usage, 13.8x could not arithmetically be mostly switching; most of it would have to be new demand.

What is s in practice? SemiAnalysis addressed it on August 18, ten days before OpenRouter's post, in connection with another promotion on the same model family. SemiAnalysis pointed out that OpenAI's 50% discount on GPT-5.6 Sol was available only on OpenRouter and on Vercel's AI Gateway, another routing layer developers use to reach models, and wrote that the two platforms, "despite being a very small portion of OAI's total token volumes," were "disproportionately impactful," because they happen to be the main public data sources the industry uses to estimate model market share (SemiAnalysis, as relayed by BigGo).

If "a very small portion" is right, s falls in the first two rows of the table, and switching alone can explain the entire 13.8x without the world using one more token.

SemiAnalysis's post contains a statement of fact and a guess at motive. The fact is that the two platforms are a very small share of OpenAI's total volume yet the industry's main public data source. The guess is that this may be clever marketing that moves the industry's visible scoreboard at low cost. This essay uses the fact and sets aside the guess. Neither OpenAI nor either platform has commented on motive, and the relay itself notes that the interpretation comes from a third-party analysis. Motive cannot be checked, but the structure can: placing a discount on a channel that is a small share of a company's own volume yet serves as the industry's scoreboard costs far less than a market-wide price cut, and the movement on the scoreboard far exceeds the movement in the actual market. That asymmetry exists without anyone needing bad intent.

The methodology chapter of OpenRouter's own hundred-trillion-token study puts this far more candidly than OpenRouter's marketing post:

"This study reflects patterns observed on a single platform, namely OpenRouter, and over a finite time window, offering only a partial view of the broader ecosystem."

And on small models:

"Small models are precisely the ones most commonly self-hosted, which means a meaningful portion of their real usage is invisible to any centralized API provider."

The research team calls its own findings "only a partial view of the broader ecosystem." The marketing post attaches the label "Jevons paradox," which is a claim about total consumption across the whole economy. The methodology chapter rules out the scope of inference that the marketing post draws.

Fifth: the reading in this episode that can carry a Jevons argument is a different one

None of this means the episode measured nothing. This episode does contain a reading that can be used to talk about demand. It is just not 13.8x.

Back to the two layers of the discount. The July 30 list-price cut was the genuinely market-wide one: Luna down 80%, for everyone, through every channel, as a standing list price rather than a promotion. That cut is the one that could produce new demand in the Jevons sense.

And the place to test it is after the discount ended. On August 14 the 50% channel discount expired, but OpenAI's 80% list-price cut stayed. From August 15 onward, Luna remained 80% cheaper for everyone than it was in early July. OpenRouter published the retention over that stretch:

"About 32% retained some usage in the subsequent days and 18% ran at or above their program pace."

(OpenRouter's explanation, cited above)

That is: about 32% of users kept some usage in the following days, and 18% ran at or above their discount-period level.

That figure is the closest thing to Jevons evidence in this whole episode, for three reasons. First, it measures behavior after the price returned to a level everyone can get. Second, it compares the same users across time, free of share-dilution effects. Third, the corresponding price change is market-wide, not a channel's.

And its size is moderate: fewer than one in five users kept up their discount-period usage. That is a completely different story from fourteenfold. In this episode, the figure that bears on how much demand a price cut sustains is the 18%, not 13.8x. The 18% is a share of users, not a share of volume.

(A limit: the 32% / 18% covers only six days after the discount, and OpenRouter notes that the post-period is six days long and the last day may be incomplete. Six days is too short to show the medium term.)

The measurement that would settle this in one step, from data that already exists

The most useful thing this essay produces is not a conclusion. It is a test that can already be run at close to zero cost.

On July 27, 28 and 29, those three days, only OpenRouter's 50% channel discount was in force; OpenAI's list price had not yet moved. Across the whole market, Luna's price did not change by a cent over those three days. Only the OpenRouter route got cheaper, by half.

So the usage change over those three days is, by construction, pure channel substitution: a response to a channel price gap, not to the market price.

• If usage had already jumped toward the tenfold range in those three days, most of the 13.8x is route switching, and this essay holds.
• If usage barely moved in those three days and only took off once the July 30 list price landed, the main driver was the market-wide cut, the Jevons reading strengthens sharply, and this essay's reservation shrinks to a footnote.

OpenRouter's post includes a daily indexed usage chart, which means OpenRouter has this data in hand. OpenRouter computed 5.6x and 13.8x as "average daily volume over the whole window against average daily volume before it," and did not break out July 27 to 29 separately. Breaking it out closes the case.

What this changes, and for whom

Enterprise technology leaders and application-layer companies. The fourteenfold figure circulating in the industry is a switching response, measured on the group most practiced at switching. That makes it a poor basis for a negotiating position or a budget assumption. The question that matters to an application company is a different one entirely: how much of its usage would disappear if its vendor raised list price by 30%. The switching cost of such a company and the switching cost of an OpenRouter user are not the same order of magnitude. No outside study has measured that number for a given company's own users. Each company has to measure it in-house.

Pricing and capacity leads at model companies. Here there is a very cheap tool and a matching risk. The tool: a discount placed on a channel that is a small share of the company's own volume yet serves as the industry's main public data source moves the scoreboard a long way for very little money. The risk: a pricing team that reads the same scoreboard will systematically overstate the revenue elasticity of its own price cuts, because the reading includes the company's own direct-API volume moving to the discounted channel. On that volume the company is paying itself, and it gives up a channel cut on the way. Before such a number can inform a pricing decision, the company's own cross-channel total has to be netted out.

Investors and data-center capacity planners. Any capacity conclusion derived as "price cut → N times the volume → so compute demand will hold up" has the wrong denominator. The same tokens taking a different route do not build another data center. Last week this newsletter asked how costly it would be for a model company to pull inference compute back for internal research. The answer depends directly on which elasticity goes in. Pulling compute back means less capacity for paying customers, and that pushes prices up. If demand reacts the way 13.8x suggests, volume collapses the moment price rises, so withdrawal is expensive. If the response is closer to the 18% retention reading, far less volume is lost, so withdrawal is far cheaper. The two answers point to opposite capital-allocation recommendations. This essay takes the latter: 13.8x measured a channel price gap and cannot be used to price the cost of withdrawing compute; the usable reference is section five's 18%, which points to a far more moderate magnitude. But that reading covers only six days, so this is a direction, not a value: the estimated cost of withdrawing compute should be revised down, and by how much is something this episode cannot measure.

The obligation in the other direction. Anyone using this essay to argue that Jevons is a fiction, or that demand for AI tokens has no elasticity, has gone just as far past the evidence. This essay has not measured the size of the substitution. It has only shown that the substitution was never subtracted, and that this experiment's design could not subtract it. Section five's 18% is positive, not zero.

What would prove this wrong

1. OpenRouter (or anyone) publishes the July 27 to 29 readings on their own, and usage barely moved over those three days → this essay's core reservation fails, and 13.8x was driven mainly by the market-wide price cut. This is the cheapest and most direct test.

2. OpenAI's share of OpenRouter's tokens before the discount, excluding Terra and Luna, turns out to be close to 27% rather than the roughly 8% the public compilations show → section three's share recalculation is void, and the post's reading that three quarters came from outside holds.

3. A total-usage reading for Luna across all channels appears, showing that total volume over the same period also grew on the order of tenfold → the substitution contamination is small, channel elasticity ≈ market elasticity, and this essay's core reservation fails.

4. OpenRouter publishes competitors' absolute usage during the window (not shares), and it shows absolute volume falling → section three's inference that competitors very likely grew is overturned.

Verdict date

2026-10-15. By then the discount window will have been closed for two months. If OpenRouter publishes its usual monthly model shares and daily data, or a third party publishes cross-channel usage, items 1, 2 and 4 above can be read off directly. This observation window is this newsletter's own, not one set by any outside body.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.