OpenAI eng lead: inference 60% faster, next model far more token-efficient

MatthewBerman · x · 2026-08-25

Highlights from Matthew Berman's long interview with Tibo (@thsottiaux), OpenAI's engineering lead:

Speed: outside ultrafast, speeds are roughly 60% faster than 3 months ago, and they will publish how. He expects ultrafast-like speeds to become the default (or close) within a year or two. Most OpenAI employees don't get ultrafast — the vast majority of capacity is reserved for customers.
Token efficiency: Sol is significantly more token-efficient than Terra, and the next model (presumably Astra) will be significantly more efficient than Sol.
Merging: ChatGPT and Codex merged because the next models are one agent that does everything.
Candid on the harness: skill files are hard to maintain, memory doesn't always remember, subagents can break the illusion.
The reset button isn't coordinated with marketing or finance — and yes, there's a physical one: "I can press the button whenever I want."

More from Companies & People