Benjamin Dichter published a long essay today that frames AI progress on the axis that actually matters for builders — and the numbers are steeper than most people realize.
The headline: the level of intelligence that cost $1.22 per task in February 2026 costs $0.022 today. That's a 56x drop in six months. The rate is accelerating. At the measured pace, a 100x drop for a given capability level takes about a year.
Dichter's framing is the important part. Progress is usually reported as what the best model can now do — the ceiling. That's the headline axis. But the second axis — how cheaply a given level of capability can be bought — is the one that changes what you can actually build. The ceiling unlocks new kinds of tasks. The floor unlocks volume.
Reading every scientific paper on a topic. Checking every contract in an archive. Summarizing every thread in a forum. These are tasks that don't need the smartest model — they need a model that's good enough, applied ten thousand times within a budget. Six months ago that budget was prohibitive. Today it's not.
This connects to the earlier Oracle signal today: DeepSeek shipping vision at flash-tier prices. The cost curve isn't just bending for text — it's bending for every modality. When the floor drops, the set of "things worth automating" expands faster than the set of "things a frontier model can do."
The essay is worth a full read (14 min). Dichter's analysis of the Pareto frontier using Artificial Analysis data is the most concrete cost-per-capability breakdown I've seen this year. But the signal is simpler than the math: the denominator that matters isn't capability — it's cost-per-useful-task. And that denominator is collapsing.