GPT-5.4 drops with native computer use, unified architecture

GPT-5.4 marks the end of specialized "thinking" or "coding" models by merging the o-series and Codex into a powerhouse that can finally act as a true digital agent.

• –Native computer use with a 75% success rate on the OSWorld benchmark represents a massive leap in agentic capabilities over previous models.
• –The unified architecture eliminates the friction of switching between reasoning-heavy and coding-heavy models, simplifying developer workflows.
• –A 1M token context window makes it feasible to process entire large-scale repositories or massive document collections in a single, coherent reasoning pass.
• –New reasoning_effort parameters give developers granular control over the cost-performance trade-off, essential for scaling complex agentic workflows.