
71.5% of the Tokens, 23% of the Bill: One Gateway's September Ledger
Open-weight models ran most of the tokens and a quarter of the spend on one production gateway. The useful lesson is about routing, not licensing.
Category
326 articles

Open-weight models ran most of the tokens and a quarter of the spend on one production gateway. The useful lesson is about routing, not licensing.

HeyGen open-sourced a renderer that turns HTML into deterministic MP4s. It cannot invent footage, and that is exactly why agents can use it.

ArtCraft is a Rust desktop studio that asks you to build the scene before you sample it. No prompt roulette, no re-rolling for a matching second shot.

Doubao Work's one-click import turns a user's history into luggage they can carry out. Office agents now compete less on model strength than on switching costs and memory.

AI performances and AI-written screenplays are out. The rule is not about tools, it is about credit, and credit is the one thing an institution can still decide.

A skill is not a setting. It is a bundle of instructions and scripts that an agent follows with whatever permissions it holds, and NVIDIA just shipped a scanner that knows 64 ways that goes wrong.

The fastest-growing repository on GitHub is a joke that happens to be correct: cut 65 percent of tokens by talking like a caveman. The bottleneck is not intelligence, it is overhead.

A model produced 400 pages of candidate reasoning. Mathematicians picked the three paragraphs that mattered. Then a machine checked the result line by line.

Git was designed around a belief about human pace. When agents write code continuously, the repository still works and the process around it stops meaning what it used to.

An always-on assistant is less like a taxi that charges while you ride and more like climate control. It is always doing something, and every one of those somethings has a price.

The payload is not code that can be read. It is a capability, compressed into weights, and that changes what defending against it even means.

A response under one second is not a spec sheet decoration. Below that line, talking to a system stops feeling like dictation and starts feeling like conversation.