Inference Prices Are Falling Faster Than Anyone's Budget Assumed
Microsoft priced its new speech recognition model at ten cents an audio hour. The previous model in the same line launched at thirty-six cents five months earlier. That’s a seventy-two percent cut in under half a year, on a product that also handles more languages and runs materially faster than it did.
Nothing about that is unusual any more, and the consistency is what makes it worth writing down. Per-token and per-unit prices across the major model providers have fallen by an order of magnitude or more since 2023, at every capability tier. Cheap models got cheaper. Frontier capability from two years ago is now priced like a commodity, because it is one.
Three forces are doing this at once. Hardware efficiency improves each generation. Inference optimisation, meaning quantisation, distillation, better serving stacks and speculative decoding, keeps finding large gains in software. Competition is unusually severe, with several well-capitalised providers plus a set of open-weight models that establish a price ceiling nobody can charge above for commodity work.
The consequences run through several layers, and they don’t all point the same way.
For anyone building on top, the immediate effect is that unit economics improve without any engineering effort. A workflow that lost money at last year’s prices makes sense at this year’s. That has been the single biggest driver of AI features shipping into ordinary software, and it explains why the same idea failed twice before succeeding. The idea didn’t change. The cost did.
The trap is planning around it. Teams that assume continued decline are making a forecast, and the decline has been steepest where competition is thickest and capability is easiest to replicate. Frontier reasoning at long context hasn’t followed the same curve, and agentic workloads consume tokens at rates that overwhelm per-token savings comfortably. Cheaper units with far more units used is not obviously a smaller bill.
For providers, this is the familiar squeeze between capital intensity and commodity pricing. Training runs cost more each cycle while the output gets priced closer to the cost of serving it. That gap has to close from somewhere: proprietary data, distribution, enterprise contracts with switching costs, or vertical integration into the silicon and the power. It won’t close from model quality alone, since quality advantages have been lasting months rather than years.
For buyers, the practical conclusion is unromantic. Sign short. Any multi-year commitment at today’s prices for commodity inference is a bet against a trend that has held for three years running. Build the abstraction layer that lets you swap providers, keep evaluations in a form that runs against any of them, and treat the model as a component rather than a platform.
The one thing this doesn’t tell you is where it stops. Serving costs have a physical floor set by silicon and electricity, and the AI power buildout is currently colliding with grid regulators who’ve started asking whether the demand is real. If power gets expensive, the floor rises.
Prices that have only ever gone one direction tend to teach the wrong lesson about which direction they can go.