Gemini 3.1 Pro feels noticeably slower than Claude models and Gemini Flash on similar tasks — long delays before responses start, worse on multi-step/agentic stuff. Anyone else seeing this, or is it a known issue?
Hi @Farman_Khan3 ,
Welcome to the community!
The Pro series models are deep thinking models and are expected to take longer to perform tasks than the Flash series. The Pro model series is meant for tasks that are more complex and may require longer reasoning. Flash series models are designed to be quick and efficient, and are suitable for everyday tasks.
So, yes, what you’re noticing is normal and expected. I hope this helps!
-TechRay
No. I don’t think that’s it. Even simple tasks are taking too much time and it’s not because of reasoning, it’s just takes too much time processing simple tasks and shows “working” it’s not even reasoning at that time just plainly stuck for a long time. Even Flash models whilst showing that they are actively analyzing files are still way slower.
This is a genuine issue and needs to be addressed because currently the Pro Tier is almost unusable.
It’s the server-side safety car when you run with SAFETY_LEVELS primed. You can confirm this by managing to disable them. I’m afraid I cannot tell you how. Also, keep in mind, if you switch models it needs to move VRAM around, which takes a while.
What’s a sidecar?
A sidecar is another program or LLM that sits server or client-side. Google has a sidecar running that runs 2.5 Lite (I believe) that checks for safety and prohibited content. It can cause massive latency.
Now, you can disable the sidecar:
- Switch to the Developer API.
- Build you own `agy1 (you’re on your own here).
What else can you do?
Not much. You can use a lock-step mode by issuing small commands at a time, which means input tokens scan faster and the sidecar moves quickly.
I wish you the best. I am working on a dynamic budgeting system that enhances Pro. I’ll release everything to GitLab when it’s ready.