Keep pulling the thread on Alex & Romain.
Designers on the OpenAI Codex team now write more code than engineers on the team did six months prior due to the use of AI tools.
OpenAI's frontier models, such as the one referred to as "GPT 5.4," are capable of analyzing or migrating codebases containing millions of lines of code.
The vast majority of the code for the OpenAI Codex application itself was generated by an AI agent.
A user was able to use an OpenAI Codex skill to create tasks in Linear, and then instruct Codex to implement and complete all of those tasks overnight.
The release of a model called "GPT 5.2 Codex" around December marked an inflection point where it became possible to delegate much longer tasks to the model.
OpenAI's Codex product experienced growth of 20x to 30x over a few months after shipping its CLI and IDE extension.
Senior engineers at OpenAI, including Peter Steinberger and Greg Brockman, now use the Codex app as their primary development tool.
OpenAI now positions Codex as the primary entry point for developers building on its platform, including for modalities like Imogen and Sora.
Peter Steinberger is helping OpenAI build the next generation of personal agents directly into ChatGPT.
OpenAI has a model named "Codex Spark" designed for rapid ideation and iteration, achieving an average speed of 1,200 tokens per second.
The OpenAI Codex team writes very few product specifications, and only does so for problems too complex for a single person to manage.
OpenAI's Codex features a "Figma skill" that can extract details, React components, and variables directly from Figma files to implement corresponding code.