AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI's Codex: A Model So Fast It Changes How You Code

· AI & I

OpenAI's Thibault Sottiaux and Andrew Ambrosino join to discuss the rapid evolution of Codex, including the new Codex app, GPT-5.3 Codex, and an even faster preview model. They describe a strategic shift beyond professional developers toward a broader audience of builders, framed around the Super Bowl ad and the app passing one million downloads in its first week. The team emphasizes that the GUI app—not a TUI or IDE fork—is the right "command center" for orchestrating multiple agents, with features like voice input, mid-turn steering, image generation, and persistent automations. A major focus is the new model's extreme speed, which they credit to a partnership with a series hardware provider (Cerebras, implied) and a rewrite of the service stack over web sockets, cutting turn latency 30-40%. They note they had to deliberately slow streaming output because text appeared too fast. Speed is positioned as transformative for in-the-flow coding, code review, end-to-end testing, and even replacing deterministic Git operations with model-driven skills. Looking forward, they identify verification—confirming that generated code actually works end-to-end—as the next major bottleneck, and outline their approach to code review and PR evidence. They also compare Codex to Claude Code and Cursor, arguing their tight integration of research, product, and engineering gives them a compounding edge.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.