In recent weeks, doing a deep research with gpt5 seems broken for me

It usually looks at less than 20 sources. Basically, it has become useless.

Conversely, just doing thinking or pro mode gives me hundreds of sources

What gives?