OpenAI researcher SQ Mah explains how GPT-5.4 Thinking brings even more powerful capabilities to Codex — with more persistent computer-use (CUA) capabilities that cut token usage by two-thirds in some cases, and stronger image understanding for seamless website UI and image generation.

YouTube

OpenAI's GPT-4.5 (referred to as GPT-5.4) Thinking model introduces two key app development capabilities: persistent computer use (KUA) and image-to-website replication. With persistent KUA, the model interacts with UIs like a human would, reducing token usage by up to 2/3 compared to previous models. A demo shows the model building and testing a 3D chess Electron app with glass and marble effects, using KUA to actually play through the game and verify functionality. A second demo shows the model converting a hand-drawn coffee shop website design into a real site, generating contextually appropriate images concurrently and using KUA to visually compare the output against the original design.

Computer Use & Frontend UI with GPT-5.4 Thinking