Cost per unit of intelligence fell by half across multiple tiers

Trackers report roughly a 50% drop in cost per intelligence unit this month, not in one tier but across several at once. A simultaneous fall means the cause is not one lab's pricing decision.

Release trackers report the cost per unit of intelligence falling roughly 50% across multiple tiers this month. The measure is rough by construction — it divides a price by a capability score, and both halves are contestable — but the direction and the simultaneity are harder to argue with.

Simultaneous is the interesting part

One provider halving its price is a commercial decision. Several tiers moving together is a cost structure changing underneath all of them: better inference efficiency, cheaper serving hardware, or smaller models reaching capability that previously required larger ones. The third is the one visible in the release list, where 27B and 30B models appear alongside frontier releases rather than well behind them.

What halving does that a 10% cut does not

A 10% price cut moves margins. A 50% cut moves the boundary of what is worth automating at all. Tasks priced out at the old number become viable at the new one, and the volume that arrives is not the old volume slightly larger — it is a different set of applications.

It also makes a category of business model precarious. Anyone whose product is a thin margin on top of inference has just watched the floor drop, which is good for their costs and bad for what they can charge.

The caveat worth keeping

"Cost per intelligence unit" is not a measured quantity. It is a ratio built from a price everyone agrees on and a capability score nobody does. Treat the halving as a real directional signal and the precise figure as an artefact of whichever benchmark sat in the denominator.

Local AI Zone — Latest AI Developments: August 2026 Update → · LLM Stats — AI Updates Today (August 2026) — Latest AI Model Releases → · AI Release Tracker — Latest AI Model Releases — August 2026 →