OpenAI researcher SQ Mah details how GPT-5.4 Thinking integrates into Codex with expanded computer-use agent capabilities and a two-thirds reduction in token consumption for certain tasks.
The gains are concrete: more persistent CUA behavior means the model sustains longer autonomous sessions without losing context, and improved image understanding lets it parse and reproduce frontend UI layouts directly from screenshots or design files. These are not incremental polish, they are architectural shifts in how the model handles multi-step visual and agentic workflows.
The video is worth watching in full for Mah's technical breakdown of where token savings actually come from and how the image pipeline connects to live website generation. The implementation details reveal tradeoffs that the headline numbers do not.
[WATCH ON YOUTUBE →]