GPT-5

GPT-5 was released on 7 August 2025, twenty-nine months after GPT-4, to every ChatGPT user at once. It unified two families: the fast GPT models and the slow reasoning o-series were replaced by a system with a router that decided, per request, whether to answer immediately or think, and for how long. It set new marks on coding, mathematics and health benchmarks, hallucinated less, and cost less per token than GPT-4o. Sam Altman compared having it to having a team of PhD-level experts on call.

The reception was the story. Users mourned the removal of GPT-4o's personality until it was restored; the launch demonstration contained a mislabelled chart; and the consensus among researchers was that the model was an excellent product and an ordinary increment, evidence that the scaling curve had bent even as the number of people using the models had reached 700 million a week. The gold-medal olympiad model of July was not the one shipped.

GPT-5 is on this timeline as the point at which the frontier laboratories' releases became quarterly and incremental, GPT-5.4, 5.5 and 5.6 following within a year, and as the moment the public conversation shifted from whether the models would keep improving to what the improvements were for.