OpenAI Cut Its Price 90%. It Made More Money.

Three models, one price experiment

On July 9, 2026, OpenAI shipped the GPT-5.6 series as three separate models instead of one: Sol, the flagship for complex reasoning and agentic work; Terra, positioned as competitive with GPT-5.5 at roughly half the cost; and Luna, the smallest and cheapest of the three. All three launched on the same day, at the same starting line, which turns out to matter a lot for what happened next.

Three weeks later, OpenAI and OpenRouter ran a pricing experiment on two of the three models and left the third one alone. OpenRouter, the API marketplace that routes billions of tokens a day across every major model provider, published the full before-and-after data trail on August 25 in a post titled “GPT 5.6 Discounts & Jevons Paradox.” We pulled the raw multipliers from that post, cross-checked the resulting dollar prices against OpenRouter's own live model pages for Terra, Luna, and Sol, and ran the one calculation the original post stopped short of: what happened to revenue, not just volume.

The discount was two cuts stacked, not one

A round number like “90% off” usually means one decision. This one was two, layered on top of each other. From July 27 through August 14, OpenRouter ran a 50% promotional discount on Terra and Luna, priced against whatever OpenAI's list price was that day. Three days into that window, on July 30, OpenAI separately cut its own list prices: Terra's input rate fell from $2.50 to $2.00 per million tokens, a 20% cut, and Luna's fell from $1.00 to $0.20, an 80% cut. Stack the two together and the blended discount from July 30 onward was 60% off Terra and 90% off Luna, the number OpenRouter used in its own headline.

Output pricing moved by the same ratios: Terra's output rate fell from $15.00 to $12.00 per million tokens, Luna's from $6.00 to $1.20. Every one of those figures is on OpenRouter's live model pages today. Sol never got a list-price cut during the Terra and Luna program. It got its own separate 50% promotional discount starting August 17, which is the detail that turns this from an interesting chart into a genuine natural experiment.

The volume didn't rise with the discount. It overshot it

If usage rose in exact proportion to the discount, a 90% price cut would produce roughly a 10x jump in volume at constant spend, since 10x the tokens at a tenth of the price is a wash. That's not what happened. Luna's daily token volume during the discount window ran 13.8x its pre-discount average. Terra, discounted less, still ran 5.6x. Sol, which OpenRouter used as the control because it kept its list price the entire time, moved from 71.2 billion tokens a day to 79.1 billion: a 1.11x bump, inside the range you'd expect from ordinary week-to-week noise on a platform routing that much traffic.

That gap between the discount size and the usage multiplier is the whole story in miniature. A 90% discount that produced exactly a 10x volume increase would leave revenue unchanged. A 90% discount that produced a 13.8x increase means something more elastic is going on: enough latent demand existed at the old price that cutting it uncorked usage faster than the discount itself grew.

Where the extra usage actually came from

The obvious skeptical read on any usage spike from a discount is that it's mostly existing customers doing the same work for less money, not new demand. OpenRouter's own share breakdown argues against that read. Before the discount, Terra and Luna combined made up 0.7% of every token routed through the platform. During the discount window, that climbed to 7.8%, an 11x jump in relative share and a gain of 7.1 percentage points.

The part that got less attention: where those 7.1 points came from. Competing labs' models lost 5.3 share points over the same stretch. OpenAI's own other models, Sol included, lost only 1.9. Do the division and roughly 3 out of every 4 tokens Terra and Luna picked up came out of a rival lab's model, not out of OpenAI cannibalizing its own lineup. Zoomed out to the whole OpenAI family, combined share rose from 7.1% to 12.4% on average across the program, and crossed 15% on individual days.

That split matters for anyone trying to model whether a price cut like this is a wash for the company running it. A discount that mostly reshuffles a lab's own product line is a coordination cost. A discount that pulls three out of four incremental tokens away from other vendors is a competitive weapon, and it's the second kind of discount that shows up in this data.

The control group that flipped sides

Every one of the numbers above depends on Sol staying a legitimate control: a model that existed the whole time, under identical market conditions, at an unchanged price. It did, for the first 19 days of the program. Then, on August 17, OpenAI gave Sol its own 50% discount. If the Terra and Luna pattern was really caused by their price cuts and not by some unrelated seasonal effect hitting the whole OpenAI GPT-5.6 family at once, Sol's volume should have jumped the moment its own discount started, and nowhere else.

That's what happened. Sol's daily volume broke its 19-day flat line on the same day its discount went live, reproducing the same jump shape Terra and Luna had shown three weeks earlier, this time on a single model with no other variable changing. A single discount producing a usage spike is a data point. A second, independent discount on a different model producing the same spike, on the day it starts, is closer to a replicated experiment, and it's the strongest piece of evidence in OpenRouter's dataset that price, not some coincidental trend, is what's driving the volume.

Who stuck around once the discount ended

Every discount produces a spike while it's live. The number that actually matters for whether it was worth running is what happens after it ends. OpenRouter tracked more than 100,000 customers who used Terra or Luna during the discount window and checked their usage in the days after the discount expired. 32% kept using one of the two models at all. 18% used it at or above the pace they'd set during the discounted period.

That 18% figure understates how much volume actually persisted, because it's a count of customers, not a count of tokens. Weighted by tokens instead, daily volume in the post-discount window ran 1.38x the program-period daily average, meaning the customers who stayed are, on average, considerably heavier users than the median customer who tried the discount and left. OpenRouter is upfront that the post-period is short, 6 days of data against a 19-day program, so this number will move as more data comes in. But the direction it's moving in, a smaller group using more, not less, is not the pattern you'd expect from a promotion that only worked while the discount lasted.

The math nobody in the coverage ran

Every number so far comes straight from OpenRouter's post. This next one doesn't. Volume multipliers and price cuts are two halves of the same equation, and multiplying them together tells you what happened to revenue, which is the number that actually determines whether a discount like this was a good trade for OpenAI. Nobody in the coverage we found ran it, so we did.

# Independent estimate, not company-confirmed.
# Effective discount and volume multiplier both pulled from
# OpenRouter's "GPT 5.6 Discounts & Jevons Paradox," Aug 25 2026.

models = {
    "Luna":  {"effective_discount": 0.90, "volume_multiplier": 13.8},
    "Terra": {"effective_discount": 0.60, "volume_multiplier": 5.6},
}

for name, m in models.items():
    price_multiplier = 1 - m["effective_discount"]          # e.g. 0.10 for Luna
    revenue_multiplier = price_multiplier * m["volume_multiplier"]
    print(f"{name}: {price_multiplier:.2f}x price * "
          f"{m['volume_multiplier']}x volume = "
          f"{revenue_multiplier:.2f}x revenue "
          f"({(revenue_multiplier - 1) * 100:+.0f}%)")

# Luna:  0.10x price * 13.8x volume = 1.38x revenue (+38%)
# Terra: 0.40x price * 5.6x volume = 2.24x revenue (+124%)

At 90% off, Luna's effective price is roughly a tenth of what it was. A tenth of the price times 13.8x the volume comes out to about 1.38x the revenue: Luna is on pace to bring in roughly 38% more money than before the cut, not less. Terra, at 40% of its old price times 5.6x the volume, comes out to roughly 2.24x, a 124% increase.

The 160-year-old theory this is actually testing

In 1865, the English economist William Stanley Jevons published The Coal Question, arguing that making steam engines more fuel-efficient wouldn't reduce Britain's coal consumption. It would increase it, because cheaper, more efficient power created new uses for power that hadn't been worth paying for before. The idea became known as the Jevons Paradox: for goods with elastic enough demand, a price cut can raise total spending on the good rather than lowering it, because the percentage increase in quantity consumed outpaces the percentage decrease in price.

Economists have argued for over a century about how far that logic travels outside of coal. Testing it directly usually requires either historical data with a lot of confounding variables or a controlled experiment nobody has an incentive to run at scale. AI API pricing turns out to be an unusually clean setting for it: OpenRouter has minute-by-minute usage logs across every model on the platform, prices are precise and change on documented dates, and, in this case, a company ran two independent discounts on near-identical products three weeks apart, producing something close to a replication rather than a single data point. The GPT-5.6 Terra and Luna discount is, as far as we've found, one of the first times this pattern has shown up with this much granular, public data behind it in the AI infrastructure market specifically.

What this doesn't prove

None of this means every AI pricing discount pays for itself. It means this one, on this data, on the assumptions above, looks like it did. A model with less latent demand at its old price, a discount that isn't layered with a second promotional push, or a customer base with less room to expand usage could all produce a volume multiplier well under the price cut instead of well over it, in which case the same math runs in reverse and the discount is a straightforward loss. The Jevons Paradox isn't a law that applies to every discount; it's a description of what happens when demand is elastic enough, and the only way to know whether that's true for a specific product is to run the experiment and look at the volume multiplier, the way OpenRouter's data lets you do here.

It's also worth being honest about what “more revenue” doesn't capture. Compute costs to serve 13.8x the token volume are not zero, and neither OpenRouter nor OpenAI has published gross margin figures for Terra or Luna at the new price. Revenue up 38% on Luna is not the same claim as profit up 38% on Luna, and nothing in the public data lets you compute the second number.

If you want to run this check on your own pricing

The two things worth knowing before you cut a price on anything metered: your own historical volume-vs-price elasticity, if you have enough data to estimate it, and a clean way to measure the before-and-after once you do cut, ideally with something like Sol in this dataset, a comparable product you didn't touch, so a volume change has something to be measured against instead of just a before-and-after on the one thing you changed. Most teams have the second piece by accident, in the form of whatever they didn't discount, and never look at it.

For AI workloads specifically, that comparison is only useful if you can actually see what you're spending per model, per call, cached and fresh tokens broken out, instead of a single blended number on an invoice at the end of the month. That's the same visibility gap this piece is really about: OpenRouter could publish this analysis because it has granular, per-model usage logs across every provider on its platform. Most teams consuming these APIs directly don't have the equivalent view into their own spend. MegaBrain is built to close that gap by default: one API across 500+ models at zero markup, with the real per-call, per-model breakdown, so a price change on any model you route through shows up as a number you can actually compute an elasticity from, not a shift in a monthly total you have to reverse-engineer. And for workloads that should be running and measuring continuously rather than only when someone remembers to check, BrainClaw, MegaBrain's always-on OpenClaw agent runtime, keeps that kind of monitoring alive in the background instead of depending on someone re-running the analysis by hand every time a provider changes a price.

Sign up at getmegabrain.com to see your own per-model spend instead of estimating someone else's.