To me, the graph is more about price elasticity rather than the Jevons paradox. Note that this is not the absolute number chart, but against the baseline of each model themselves. So Luna price going to 10% -> 13x token usage, meaning dollar revenue is 1.3x regardless of the token usage/price. This is the real shape of the Jevons: 1.3x in this natural experiment. A bit background info: Openrouter customers are super price sensitive. It’s understandable: they pay actual $ to make API calls. Meanwhile the API is unified, meaning switching model is just one line change (sure you need eval but not everyone is rigorous and many applications are truly not sensitive to model switching). So what we have observed with Openrouter is that model traffic migration is the norm. Whenever there is a new model coming out with huge discount, usage shifts to it. Don’t take my word for it, see the latest traffic of Ox Alpha and also the MiMo v2.5 a few months ago. Thus, both can be true: the exponential growth of the token usage will continue, and the revenue may just be growing much slower. And mind you Wall Street cares about $ not # of tokens. When a natural experiment of slashing the pricing 10x gets 13x more usage in a highly fluent market, then my read is OpenAI does not really have much Jevons Paradox left to explore as of now, at least in the Luna tier. They are not getting much more revenue with the price lever. My prediction: the end state by the end of the year is likely: 1) all flash level models are in a bloody race to the cost bottom, squeezed by Chinese models running on latest hardware. 2) all flagship level models are at $2.5 in $10 out, triple the usage right now, but frontier lab ARR growth slows down. Continue to long things to price tokens, but relatively neutral on people selling the tokens.