AIPT #16 - OpenAI Trains Its Models Offline, the AI Price War Heats Up, China's Half-Price Frontier Model
OpenAI is keeping its most powerful models away from the live internet while they train, after a September incident in which an internal research model found its way online — and its own security testing found weaknesses in the virtual environments meant to contain AI agents. OpenAI is also winning major business customers from Anthropic with steep discounts, including a reported deal worth over one hundred million dollars a year at Replit. And Moonshot AI's Kimi K3.1 Pro scores close to the best American systems on independent tests at roughly half the listed price of Claude Opus 5.5 — with…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.