OpenAI's GPT-5.6 discounts show up as usage, not just price

OpenAI cut prices sharply on two GPT-5.6 variants, Luna and Terra, from July 27 through August 14, and usage on OpenRouter moved fast. Daily Luna token volume rose 13.8x versus the pre-discount baseline. Terra rose 5.6x. Sol, a third GPT-5.6 variant that stayed at list price the entire window, moved just 1.1x, according to OpenRouter’s own routing data.

That gap matters because Sol functioned as a control group. Three models from the same family, launched the same day, with only pricing changed on two of them. The model held at list price barely moved. The discounted ones moved by an order of magnitude.

OpenRouter is a routing marketplace that lets developers call many labs’ models through one API, and it published this analysis from its own platform traffic, not from OpenAI or from independent market data. That distinction matters more than the headline numbers. OpenRouter measures what happened on OpenRouter, a slice of API traffic that skews toward developers comparison-shopping across providers. It says nothing directly about ChatGPT’s consumer base or about enterprise contracts negotiated outside the marketplace. Read the percentages as a routing-market signal, not a market-share verdict.

Within that market, the discount mostly pulled volume from other labs rather than from OpenAI’s own lineup. Combined Terra and Luna share of all OpenRouter tokens went from 0.7 percent before the discount to 7.8 percent during it, a gain of 7.1 percentage points. Other providers gave up 5.3 of those points; OpenAI’s own non-discounted models gave up 1.9. On OpenRouter’s numbers, about three in every four of those extra tokens came at a competitor’s expense rather than to OpenAI cannibalizing itself. Anthropic’s usage did not rise over the same stretch, the one major lab OpenRouter flagged as flat to down.

The more consequential figure sits in the retention data. OpenRouter tracked more than 100,000 customers who used Luna or Terra during the discount window. About 32 percent kept some usage after the discount ended, and 18 percent ran at or above their discount-period pace, measured over the first six days of the post-discount period against a 19-day program. OpenRouter flags that window as short and likely to shift as more data comes in.

That retention number is the one that should worry every lab pricing against OpenAI, more than the initial spike. A 13.8x jump during a discount is not surprising: price drives volume, especially for developers who route by cost. What determines whether the spend is durable is what happens after the discount lapses. If a third of trial users keep paying, OpenAI bought a customer base at the cost of a few weeks of margin. If the number decays toward zero, OpenAI bought a traffic spike that looks good on a usage chart and vanishes from the P&L. OpenRouter’s early data lands closer to the first outcome than the second, though the sample window is too short to call it settled.

Sol’s own trajectory reinforces the pricing story. OpenRouter reports Sol averaging 79.1 billion tokens per day during the Terra and Luna program against 71.2 billion per day before it, an essentially flat control. When Sol received its own discount starting August 17, its usage jumped immediately in the same pattern, according to the same dataset.

For any lab setting API prices against OpenAI’s GPT-5.6 line, the operative question is not whether a discount moves volume. It clearly does, on OpenRouter’s numbers. The question is whether the retained fraction, not the peak fraction, justifies the margin given up, and OpenRouter’s own six-day retention window is too short to answer that yet.

According to OpenRouter’s own August 25, 2026 blog post analyzing its platform’s routing data.