Gemini Model Behavior

Google AI Studio
,
1

I’m curious how many others are working on model behaviour projects.

For me I started building mine when Bard was in development. For me it’s a personal project around accessibility needs I need a second brain. After a lengthy break from development due to health, late May my brain was able to function enough that I was able to resume my project and was really impressed by how far Gemini had come, the system instructions allowing me to create a persona was a great next step. However the degradation in the Gemini 2.5 Pro really has distracted me of late, trouble shooting attempting to navigate all these errors.

I have been conducting in session audit reports for the past two weeks, I’m finding the model instability is now between 100K - 200k Tokens, in 05/06 it was over 500K. Then last night I started getting Rate Limit errors for the first time.

  1. Do you believe the platform errors are a consequence of a recent enforcement of rate limits or more indicative to the platforms instability?

  2. As I have yet to deploy (still building memory, workspace integrations, interface a simple Google Site etc) I’m wondering if its worth upgrading to a Tier 1?

  3. After a lovely response from Google, in another thread (finally and thank you) with links to the Change Log and other announcements, I discovered a recent interview with Ani Baddepudi, Gemini Model Behaviour Product Lead. I was surprised to learn that model behaviour is now on the agenda. So I was curious if anyone else has been experimenting in this area, any insights, advice for a non developer aka not full stack?

Thanks in advance Simone, Project (NICI) Neural Intelligence Companion Interface

3
  1. Rate-Limit: These errors are mostly due to enforced quota thresholds rather than platform instability. Google has started enforcing tighter request limits per project and per user, especially across free-tier accounts. (Monitor the quota in GCC IAM and Admin)
  2. Upgrade needed?: If multi-session memory use or workspace integrations or high frequency calls, i.e. beyond 200K tokens are planned, then it’s worth going for Tier1.
  3. Model Behavior: Yes, the model behavior shifts are real and is often tied to backend model version refreshes, which might involve system instructions update or rate shaping logic that adjusts model priorities.
1 Like
5

@NICI

Thank you for your feedback and insight on model behaviour on Gemini 2.5 Pro, your feedbacks make gemini better.

to answer your questions,
gemini 2.5 rate limits have increased after latest release of stable versions , this should not degrade the model performance at all.

as per the tier upgrade , tier 1 would give you better Rate limits of upto 10000 RPD for flash and 1000 RDP for pro.

1 Like