GPT-5.4 Released: Three Variants, Native Computer Control, 33% Fewer Hallucinations

OpenAI launched GPT-5.4 on March 5th in three variants: Standard (balanced), Thinking (reasoning-first), and Pro (maximum capability). The unified model combines advanced reasoning, native coding support, and direct computer interaction into a single system. Benchmarks show 75% accuracy on OSWorld (surpassing human baseline of 72.4%), 87.3% spreadsheet accuracy, and 57.7% on SWE-Bench Pro. Regional processing endpoints incur a 10% upcharge, reflecting enterprise data sovereignty demands.