When using GPT 5 in the Assistant, the model frequently (>50% of the time) fails to retrieve the search results provided by the ResearchAgent. You can see in these thinking traces that the model seems to have gotten blank results from the tool call, or something along that line.
Sometimes it works as expected though:
And links to the chats where this occurred:
https://kagi.com/assistant/b6e32ee8-7351-4635-bef6-24f95e7b857d
https://kagi.com/assistant/a91e8065-e2df-4b89-a5a6-85bb2143f934
https://kagi.com/assistant/63900f6b-436b-4be1-bc3b-062663b355f8
And a chat where it seemed to work:
https://kagi.com/assistant/63fd8b1c-aa22-4e60-933f-1ac165cb1aca
I expect the model to respond similarly to the last image I included, going through search results instead of being confused about there not being search results.
EDIT: It seems that search actually doesn’t work at all. Nearly all citations it provides are hallucinated. It also seems to affect all GPT 5 models (mini and nano included).
