GPT-4-class intelligence got 938× cheaper in three years. Intelligence is deflating faster than any technology input in history.
I built a pipeline on Bonacci Studio that snapshots live AI pricing, capability benchmarks, and GPU rental costs, then asks a simple question: if you hold capability fixed, what does it cost over time?
Hold Capability Fixed
Hold it at GPT-4's intelligence — its 2023 benchmark score — and track the cheapest model that still clears that bar:
The capability that defined the frontier in 2023 is now served by a 2-billion-parameter model at four cents per million tokens.
Three Things the Data Made Sharper
- The collapse is uneven. The frontier holds its price ($3.40/M median); the efficient tier is where the floor falls out ($0.45/M). Buying the best model has never been cheap. Buying last year's best has never been cheaper.
- Price is now a property of the seller, not the model. The same open model (DeepSeek V3.2) is served across 12 providers at a 13× price spread. That's commoditization happening in real time.
- The API price war has out-competed your own hardware. Convert Vast.ai GPU rents into $/token and self-hosting open weights costs 9× to 54× the cheapest API. You'd need impossible (>100%) GPU utilization to break even. Inference is now cheaper than the silicon.
The Pipeline That Keeps Running
The interesting part isn't the snapshot — it's that the pipeline keeps running. Scheduled ingestion, change detection that flags the moment a provider cuts a price, an append-only store that becomes a proprietary dataset the longer it runs. History you can't buy after the fact.
Intelligence is deflating ~10× a year at constant capability. The question for anyone building on it isn't "is it getting cheaper" — it's "am I capturing the deflation, or pricing as if it stopped."