AI models keep getting cheaper. Why is that scary for the labs?
Look at the latest price tag. OpenAI's new model for operating a computer, GPT-6.1 Terra, costs 1 dollar per million input tokens and 5 for output. That is half of what the previous model in its class cost. The company also says it matches a far more expensive flagship on a hard agent benchmark at roughly a tenth of the token price, and runs more than three times faster. Those are the company's own numbers, so take them with a grain of salt.
Good for customers, obviously. Now flip the chair and sit on the other side.
If prices keep falling and the models keep improving, who exactly is making the money?
That is the squeeze. The costs on the lab side do not fall. Anthropic's own filing describes compute commitments in the hundreds of billions, mostly impossible to cancel, and it names cheaper open-weight rivals as a threat to pricing. Chips are increasingly bought on debt; one space and AI company just locked in 40 billion dollars of GPU financing. When the price per unit falls but the bill is fixed, you need volume to explode just to stand still.
Is the bubble about to pop?
Nobody can promise that, and anyone who does is selling something. What we can say is what has to be true for it not to pop. Demand has to grow faster than prices fall. There is a real argument that it will: cheaper tokens make agents affordable, and agents burn far more tokens than chatbots, so total spend could still climb.
But look at the other side of the scoreboard. A chief economist at a large investment firm points out that outside the biggest tech companies, revenue per worker and profit margins have been flat, with no visible AI payoff yet. And the Chinese AI stock rally has already reversed hard.
So the question for the next year is not whether models get better. They will. It is whether the world pays enough for better to cover the cost of building it.