Aloha 
Even with the latest LLM, I can have a few minutes of generation response time despite my agent not having documents in the Knowledge Base, weak reasoning, and a light system prompt. It can also have answer in just a few seconds
I noticed the difference is when the model performs a web search. I understand searching the web adds generation time, but the delay is quite noticeable.
In the workflow agent I see gpt-4.1.mini is in the loop
I wonder if our agent lose performance at that stage of the process ?
I see on Openrouter gpt 4.1mini is at an average of 30 token/seconds and didn’t perform well at current criteria, it’s an 16 months old model.
So, is it possible to have a new Web Search action where the sub-call is doing by another model than 4.1mini ? To see the difference.
Latency is a recurrent topic and a challenge for Pickaxe and I have the conviction that the problem is here. Latency is only when a websearch is made, Sonar who have a native websearch tool so don’t use this subcall with 4.1mini have a fews seconds generating response.
I see on Openrouter GPT 5.6 Luna is less expensive than 4.1mini.
Search the web is so important we will all benefit from having a modernized tool 
What do you think of that ?