GPT-5 launch week
The model finally shipped. A same-week note.
GPT-5 launched. Two and a half years after GPT-4. The wait was long enough that the name had almost become a joke.
A few notes:
The pitch is unification, not a leap. GPT-5 isn’t a single model so much as a router. You send a prompt, the system decides whether it’s a “think fast” kind of question or a “reason hard” kind of question, and it dispatches to the appropriate underlying model. The user experience is that you no longer pick between o-series reasoning models and 4o — there’s just one endpoint, and it does the right thing. This is a real UX improvement and also a way to present incremental model improvements as a bigger jump than they are.
The benchmarks are better but not shocking. SWE-bench, GPQA, AIME — all up, some meaningfully. None of the numbers are a 4-to-5 jump the way GPT-3 to GPT-4 was. The people expecting a capability discontinuity didn’t get one. The people expecting a capability plateau also didn’t get one; the curve is still bending, just more gradually.
The rollout was rough. The first 24 hours had a wave of complaints that GPT-5 felt “worse” than 4o for casual chat, because the router was sending short prompts to the cheaper underlying model and the personality was different. OpenAI ended up restoring 4o as an option for paid users within a couple days. Worth remembering: model releases now have UX regressions, not just capability wins, and people get attached to specific model personalities in a way that matters.
The competition caught up faster than the gap. When GPT-4 launched, nothing else was close. When GPT-5 launched, Claude Opus 4.1 and Gemini 2.5 Pro are right there, and the open-weights models (DeepSeek, Llama) are a year behind instead of three. The frontier is crowded now. That’s new and probably permanent.
What I want to note for later: the era of one lab being clearly ahead ended sometime in the last 18 months, and GPT-5’s launch is the moment that became visible to everyone, not just people who follow this closely. OpenAI is still in the lead. They’re no longer alone.