Gemini 2.5 Flash latency increase after June 15 — anyone else seeing this?

Gemini API
,
1

Hi everyone,

I’m trying to understand whether other teams have noticed a recent latency degradation with Gemini 2.5 Flash on Vertex AI

In our production workloads, Gemini 2.5 Flash used to respond very quickly for our context sizes. Some prompts that previously completed in around 6 seconds are now taking around 20 seconds, and in cases with larger input token counts we are often seeing responses take more than 3 minutes. This is directly impacting users on our platform.

The timing seems notable: I started observing this instability right after June 15, 2026, which is the same date mentioned in Google’s email about access policy changes for models such as gemini-2.5-flash, gemini-2.5-flash-lite, and gemini-3-flash-preview.

I’m not sure whether this is just a coincidence, but I’d like to know:

  1. Has anyone else observed similar latency increases with Gemini 2.5 Flash after June 15?

  2. Is there any indication that Google reduced capacity or changed serving behavior for this model?

  3. Are there recommended mitigations other than migrating to a newer model?

For our use case, moving to Gemini 3.5 Flash would create a significant cost increase, while Gemini 3.1 Flash Lite performed much worse in our internal task validations.

I really like the cost-benefit ratio of Gemini 2.5 Flash, and I would prefer not to migrate away from this model or move workloads out of GCP unless absolutely necessary.

2

Yes, we’re seeing a similar issue.

In our case, Gemini 2.5 Flash with Web Grounding for Enterprise has recently become noticeably slower. Requests that previously completed much faster are now taking significantly longer, especially for complex prompts.

As a temporary mitigation, we moved to Gemini 2.5 Flash Lite. This improved latency, but we’re seeing a noticeable drop in output quality, particularly for complex search questions.

So from our side, the latency degradation does appear real, but Flash Lite is not a fully equivalent replacement for quality-sensitive Web Grounding for Enterprise use cases.