OpenAI drops GPT-5.5 with agentic reasoning

GPT-5.5 marks the transition from chat-based assistants to autonomous agents capable of managing "messy" real-world workflows with minimal supervision. The model introduces a "Thinking" mode that utilizes internal compute to verify assumptions, significantly reducing reasoning hallucinations in high-stakes tasks. It scored 82.7% on Terminal-Bench 2.0, demonstrating superior command-line and systems reasoning for DevOps and engineering, and resolved 58.6% of real-world GitHub issues on SWE-Bench Pro in a single pass. Served on NVIDIA GB200 infrastructure to maintain GPT-5.4 latency while increasing intelligence and token efficiency, the release also includes high-stakes safeguards and a specialized "Pro" tier, signaling a strategic shift toward enterprise-grade reliability over consumer chat.