GPT-5.4: OpenAI's First Model That Uses Computers Better Than Humans

OpenAI released GPT-5.4 on March 5, 2026 — the first general-purpose model with native computer-use capabilities that exceed human performance on desktop tasks. With a 1M-token context window, 75% on OSWorld-Verified (surpassing the 72.4% human baseline), tool search that cuts agent token costs in half, and three variants (base, Thinking, Pro), GPT-5.4 represents a concrete step toward autonomous AI agents that can operate software without APIs. This analysis covers the architecture, benchmarks, pricing, Codex integration, competitive positioning against Claude Opus 4.6 and Gemini 3.1 Pro, and the honest limitations including an 89% hallucination-instead-of-admitting-ignorance rate.