The Agentic Era · Product · 2025
GPT-5
GPT-5 hid the model picker. Deciding how hard to think about a question became the system's job rather than the user's.
OpenAI released GPT-5 on August 7, 2025. The most consequential thing about it was structural rather than numerical. GPT-5 was not a single model but a unified system: a fast, efficient model handling most queries, a deeper reasoning model called GPT-5 thinking for harder problems, and a real-time router deciding between them based on the complexity of the conversation, the tools required, and the user's stated intent.
GPT-5 marked the point where inference-time compute stopped being an exposed setting and became an internal allocation decision. Since o1, the field had understood that letting a model think longer improved hard-problem accuracy, but that capability had been surfaced as a choice for the user to make. Routing turned a dial into a policy. It also changed the economics quietly, because a system that decides how much computation each request deserves can serve a free tier at frontier quality without spending frontier compute on every question.