The falling cost of AI just proved it's not a commodity—it's electricity in 1910, and everyone just realized they can light up the whole factory.
The Summary
- OpenAI cut GPT-5.6 Luna prices 80% and Terra prices 20%—usage exploded 14x and 5x respectively, driving revenue UP 34% and 45%
- This is textbook Jevons Paradox: making a resource cheaper unlocks exponential new use cases, not just more of the same old ones
- The margin structure of AI just flipped—volume beats price, which means the infrastructure layer is about to get very crowded
The Signal
TD Cowen analysts tracked OpenRouter usage data and found something that should make every AI company rethink their pricing model. When Luna's effective price dropped 10x, consumption didn't just double or triple. It went up 14x. That's not customers doing more of what they were already doing. That's new behavior entirely.
This is the moment when AI stops being a tool you budget for and becomes infrastructure you assume. When electricity got cheap enough, factories didn't just run their existing machines longer. They redesigned entire production lines around abundant power. When bandwidth got cheap, we didn't just download faster. We invented streaming, cloud gaming, and always-on video calls.
"When you slash a product's price by 80%, it's pretty unusual to actually generate more revenue."
The Terra model tells the same story at a smaller scale. Price down 3x, usage up 5x, revenue up 45%. These aren't edge cases. They're proof that AI demand is elastic in ways traditional software never was. Every price cut doesn't just serve existing customers better. It activates dormant use cases that were sitting just above the economic threshold.
What this means for the agent economy:
- Agents that were too expensive to run continuously become viable as always-on services
- Tasks that couldn't justify API costs—like real-time analysis of low-value data streams—suddenly make sense
- The "good enough" tier (Terra) saw 5x usage growth, suggesting price-sensitive developers are building things they couldn't afford to test before
What this means for infrastructure:
- If usage grows faster than prices fall, total compute demand is about to spike
- Nvidia's datacenter revenue becomes more predictable, not less, as AI commoditizes
- The bottleneck shifts from "who can afford AI" to "who can deliver it fast enough"
The Implication
Watch what gets built in the next six months. The projects launching now were economically impossible at old pricing. That's your signal for where the agent economy is actually headed, not where the pitch decks say it's going. If you're building in this space, the question isn't whether to drop prices. It's whether you can scale infrastructure fast enough to meet the demand that cheap AI unlocks.
This is also a warning shot to anyone betting on AI margins staying high. The race to the bottom just proved it's actually a race to ubiquity. The winners won't be the ones who protect pricing. They'll be the ones who can serve 10x the volume at 10% the cost and still keep the lights on.