I plugged GPT-5 API into our AI products the day it launched, then excluded it behind a feature flag because the responses felt slow and the quality underwhelming. Then I read the AMA and it turned out the service hadn't fully died, but it had suffered a serious incident. Even the so-called chart crime was apparently a fatigue issue...
-
"On the day GPT-5 first launched, severe technical problems occurred unexpectedly. In particular, the automatic model switcher was down for most of the day, which made GPT-5 feel much dumber than it actually was."
-
"During the live presentation, a bad data visualization was shown because of accumulated fatigue, but the official blog post and system card later published the correct data."
-
Many users strongly asked for GPT-4o to return, and OpenAI responded by allowing Plus users to use GPT-4o again.
-
The GPT-5 team said they are aware that recent over-filtering around life sciences research, including gene therapy and bioengineering, ended up blocking legitimate academic work, and that they are working on improvements while planning specialized researcher access for lawful use cases.
-
On pricing, they said they are considering introducing a new tier between the current $20 Plus and $200 Pro. They are also exploring UI improvements to make it easier to enter "thinking mode" and support more personalized instructions.
-
Codex CLI now supports GPT-5 for paid ChatGPT users. For Pro users, the goal is effectively an almost unlimited experience. Plus and Team users can do several long sessions per week, roughly 1 to 2 hours each. Rate limits reset every five hours, with an additional weekly quota.
-
They also shared some funny internal lore, like the model's strange preference for purple-ish design choices or its habit of repeating phrases like "let's build this with care together," which has apparently become a team meme.
-
They wanted to push context windows to one million tokens, but because of cost and GPU constraints they are currently at around 128k for Pro and 32k for Plus. Long term, they want to unify reasoning and non-reasoning models and move toward token-based billing and limits.
-
Their longer-term vision includes stronger voice interfaces and eventually even brain-interface driven, "screenless" access, so users can spend less time staring at devices while still getting maximum value from the service.
