The model picker came back
GPT-5 replaced ChatGPT’s model picker with an automatic router, and the picker was back within a week. What that says about automatic choices.
When OpenAI launched GPT-5 on August 7, it replaced ChatGPT’s model picker with a router that chooses for you. Within a week the picker was back. The router wasn’t a bad idea. Taking away the override was.
What happened to ChatGPT’s model picker?
GPT-5 shipped as one system with a real-time router that decides whether a question gets a quick answer or more time to think. In ChatGPT, people could no longer choose GPT-4o, o3 or the other older models. There was no deprecation period, as Simon Willison pointed out the next day.
Then the router broke. On August 8, Sam Altman said the autoswitcher had been out for part of launch day, which made GPT-5 seem far less capable than it was. By August 11, GPT-4o was back behind a setting, and Altman called the sudden removal a mistake. On August 12 the picker returned with three choices: Auto, which uses the router, and Fast and Thinking, which skip it.
Why people wanted the picker back
The old picker was a mess: a list of names like 4o, o3 and o4-mini that only enthusiasts could tell apart. Replacing it with one model that decides how hard to think is the right instinct. Most people shouldn’t have to know which model they need.
But some people did know, and they’d built habits around it. Some picked the slow model for hard problems. Some liked how a particular model talked. When the router chose differently, they had no way to see why and no way to say “not that one.”
A default can replace a choice people made every day. It can’t remove it.
Automatic needs a rule and an override
At Uber, the experimentation platform could roll a launch back on its own, and we were careful about how. The automatic rule was written out where everyone could read it: which metric, how far past control, for how long. It acted on all traffic only. Anything narrower stayed a person’s call.
Two things made that safe to automate. People could read the rule, so the system’s choices were never a surprise. And a person could always step in and act differently.
The returned picker restores the second: Auto by default, with Fast and Thinking for people who want to choose. It’s still thin on the first. Auto doesn’t tell you why it answered quickly when you wanted it to think.
That’s the pattern for any automatic decision in a product, from a router to a smart default to an agent picking a tool. Make the automatic path the default. Say what it chose and why, in a line. Keep the manual path one tap away for the people who know better.