I’m seeing a massive slowdown after enabling web_search with GPT-5 Responses API:
Without web_search: 1-2 minutes in “in_progress”
With web_search: 40+ minutes in “in_progress”
Setup:
Model: gpt-5
background=True
enable_web_search: true
reasoning_effort: "medium"
Complex research prompts.
Question: Is this 20-40x slowdown expected when combining GPT-5 reasoning + web_search? Any optimization tips?
Output quality is great, but 40+ minutes is challenging for production use.
Additionally, the previous tests were run during office hours (PST). Now that I am running them after hours, the completion time appears much better:
in_progress (elapsed: 5.6 minutes)
status: completed (elapsed: 6.1 minutes)
Is this to be expected? Are others experiencing similar differences based on time of day, or was today just a fluke?
next time you may check here… there has been some issues yesterday afaik
Great idea. I’ll check the status page. A 40+ minute wait is just unacceptable, and having analysts run the pipeline only during off hours isn’t a viable solution.
A crawler for websites to fill multiple ontologies in a graphdb as a fallback isn’t that hard to build as well.
That’s an interesting idea! I hadn’t really considered that approach before. My use case might be tricky for pre-crawling though, I think???. I have about 10+ prompt templates. The templates don’t change, but the assignments the agent creates do, since the company changes each time. So the questions end up being pretty variable.
I’ll keep an eye on the status page and see if that pattern keeps up. I appreciate the suggestion, and it’s definitely something to think (even re-architect) about if this keeps being an issue.
Just send me a message in case you need help on that. I do that kind of stuff just for the fun of it in a couple hours.
Will do!!.
I’ve decided I might just write a custom function tool to call SerpAPI instead of using the native web_search and see if the output and latency is better