Having issues all the day in the assistant API, it takes 20-30 sec to generate a short 1 sentence answer (that’s breaking all our project)
Any ideas how to optimise it? Or when the OpenAI servers will come back to normal capacity?
(the same assistant before answered in 3-6 sec)
That’s totally unreliable…
I’m having the same problem. Just today when I’m showcasing our product to clients and it’s dead slow.
Runs stay queued for several seconds.
In API it’s even slower than in Playground… and we are getting server errors in all the workflow after because of this answer delay…
Welcome (back) to the dev community!
Did you increase your prompt? What Usage Tier are you?
Might just be network congestion…
Today I made my prompt almost 30% less, but it didn’t affect the speed at all.
Internet connection affecting the API answer speed? First time hearing that
Not internet per se, but their datacenters being bogged down trying to make sure all requests go through.
Under a minute is still good, imho, all things considered!
Good luck with it.
Under a minute is not good in our case, when it was 5 sec.
My tier is 3 or 4, I don’t remember exactly
There’s always the option of converting your project to Completions. My bot uses it’s own chain of thought loop and uses only Completions. Answers using smaller models even involving functions are almost always sub 5 seconds or less. And humans probably don’t need a response in less than one second 
For example, see this 2 second response:
Same issue. Last week was not doing file search consistently and I forced file search in the API call which seemed to fix it. But now responses are so slow, it’s just not usable.
Same here guys. Assistants API has been way too slow lately. Have you reported this to OpenAI support already?
Assistants API is junk. It’s been documented here by myself and many others. Use Chat Completions, ideally in a wrapper that allows you to switch to another provider when needed.
Just tested the API; normal responses experienced increased delays, but the tool calls had a huge delay. 
No matter how they change the names of the models, they are all the same—experiencing the same problems again and again.
They will fail, no matter how big the model is, because it is just predictive and not truly intelligent. Increasing a model’s size typically increases computational costs because larger vectors used during inference require more resources, which often leads to improved prediction accuracy (to make people believe they have reached AGI, though the bubble will burst in the coming months).
This format, where they control the model from the backend, will never work and is not reliable.