GPT-5.4 jumps in computer use, GDPval

OpenAI finally seems to be shipping a model where the interesting story is agent performance on real tasks, not just a generic “smarter than before” claim.

• –GPT-5.4 is being pitched on GDPval and OSWorld-Verified gains, a stronger signal for real-world automation than a generic benchmark bump
• –Native computer use matters for Codex-style agents and UI automation because it turns frontier model progress into directly usable workflow progress
• –The 1M-token context window is a headline feature, but long-context cost and quality tradeoffs will still matter in production
• –For AI developers, the bigger story is OpenAI collapsing chat, coding, search, and computer use into one frontier model family