OpenRouter Data: GPT-5.6 Discounts Drove 13.8x Luna Token Surge — 32% of Users Stuck Around
OpenRouter published a detailed analysis of what happened when OpenAI discounted GPT-5.6 Terra and Luna on its platform from July 27 to August 14. The numbers illustrate a pricing dynamic the AI industry has largely avoided examining directly: cheaper frontier models do not merely transfer demand, they generate it.
The Spike
During the 19-day discount window, daily Terra token usage on OpenRouter rose 5.6x compared to the pre-period average. Daily Luna token usage jumped 13.8x. GPT-5.6 Sol, which remained at list price throughout (until August 17, when it received its own 50% discount), rose only 1.1x — a useful control confirming the effect was price-driven.
OpenAI’s Terra and Luna discounts went deeper than the headline figures. On July 30, OpenAI cut its own list prices: Luna 80%, Terra 20%. Stacked with the OpenRouter discount already in place, the effective discount from July 30 became Luna 90%, Terra 60%.
The token explosion was not incremental. The pre-period had Terra and Luna combined at 0.7% of all OpenRouter tokens. During the program, combined share reached 7.8% — a gain of 7.1 percentage points. OpenAI’s total family share (Terra + Luna + Sol + other OpenAI) grew from 7.1% to 12.4%, occasionally cresting 15% on peak days.
Where the Share Came From
Three-quarters of the gain came from non-OpenAI models. Competitors gave up 5.3 share points; other OpenAI models gave up 1.9. The implication: the discounts did not primarily cannibalize GPT-5.4 or GPT-5.5 users. They brought in developers and companies running on Anthropic, Google, and other providers who switched when the price gap widened.
Anthropic specifically did not see token growth during the Terra/Luna program window, making it one of the few providers to lose share without a corresponding discount of its own.
Retention
The more consequential number for OpenAI’s long-term strategy is retention. Of the 100,000+ customers who used Terra or Luna during the discount period, approximately 32% kept using one of those models in the days after discounts ended. 18% ran at or above their discount-period pace.
Weighted by tokens rather than customer count, the picture is more striking: post-program daily volume averaged 1.38x the in-program daily average — meaning the customers who stayed were significantly larger than the median discount-window user.
OpenRouter notes this is a short window (6 days post-program vs. 19 days in-program) and the story may shift.
The Jevons Pattern
The classic Jevons paradox: when a resource becomes more efficient or cheaper, consumption rises enough to offset or exceed the savings. In AI inference, this plays out differently than in energy. Cheaper tokens don’t just mean the same workload costs less — they enable workloads that were previously cost-prohibitive. Developers running 100 API calls per user session at $0.15 per session had a different economic case when that dropped to $0.015.
OpenAI appears to have calculated this deliberately. The discount window is best read as a user acquisition program, not a margin concession. The 13.8x Luna token surge suggests it worked. Whether 32% retention is the right number to sustain a market-share shift at full price is a separate question the data cannot yet answer.
The Sol 50% discount on August 17 — after the Terra/Luna window closed — suggests OpenAI is rolling the approach up-market. Sol at 50% off targets the customer segment that ran the pre-period control: heavy, high-quality inference workloads that were not moved by Terra and Luna pricing alone.