We are getting a lot of 429 Errors calling the Vertex API, (paid) Priority tier is not respected

Google Cloud Build with AI Generative AI & Foundational Models
1

Hi all,

I’m hitting a high rate of 429 RESOURCE_EXHAUSTED (“Resource exhausted. Please try again later.”) calling gemini-3-flash-preview on Vertex AI’s global endpoint (streamGenerateContent). Roughly 40% of requests fail, and it happens even with a single request in flight — not just under concurrency. The responses include no Retry-After, and I already retry with exponential backoff.

I’ve tried priority mode to get around it — setting serviceTier: SERVICE_TIER_PRIORITY plus the X-Vertex-AI-LLM-Shared-Request-Type: priority header — but I still get the same 429s, so it doesn’t seem to help.

Questions:

  1. Are preview models on the global endpoint subject to tight shared quota that isn’t increasable until GA?
  2. Is the 429-at-near-idle expected behaviour for shared/dynamic quota?
  3. Does the priority tier need separate provisioning to actually take effect — or is it not supported for preview models?
  4. Is the recommended fix to switch to a GA model (e.g. gemini-2.5-flash) or to use Provisioned Throughput?

Thanks!

2

same issue

3

Same here, any news on this ? Did you guys find a workaround ?
Even X-Vertex-AI-LLM-Shared-Request-Type: priority header does nothing

4

Same issue here

5

Facing the exact same issue here with Veo 3.1 on my paid account (Project ID: [PII Removed by Staff]). Billing is active and I have zero usage, but I am constantly getting the ‘Quota Exceeded / 429 Resource Exhausted’ error. Please look into this backend sync bug.

6

Same here!

7

We are experiencing this heavily in the last few hours. Are others experiencing it as well today?

8

Yes, starting between 9 and 10am PT we started receiving 429 errors from the Vertex API (might now be called the Agent Platform?).

In previous hours we had many more invocations per second and no 429 errors.

9

Does google usually open incidents when this happens, or is it just resolve on its own after time.

Can’t find much online.

10

I’m struggling to figure out how widespread this is. We switched to a Pro model instead of Flash which isn’t getting rate limited for us.

I haven’t had great success with Google acknowledging these sort of capacity outages, except to say “buy provisioned throughput”

11

They just posted an incident for the 429s. Thank goodness.

12

that’s some good news – do you have a link?

13

https://console.cloud.google.com/servicehealth/incidentDetails/.

Screenshot 2026-07-16 at 3.32.15 PM
Screenshot 2026-07-16 at 3.32.15 PM1178×1216 123 KB

14

I have had the same issue - priority doesn’t appear to have any effect. In fact even using Google’s toy example from the PayGo priority docs doesn’t lead to a priority response. If anyone has any guidance on how to achieve priority requests it would be much appreciated.

Example of priority headers not working:

curl -X POST
-H “Authorization: Bearer $(gcloud auth print-access-token)”
-H “Content-Type: application/json; charset=utf-8”
-H “X-Vertex-AI-LLM-Request-Type: shared”
-H “X-Vertex-AI-LLM-Shared-Request-Type: priority”
“https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-2.5-flash:generateContent” -d
$‘{
“contents”: {
“role”: “user”,
“parts”: { “text”: “PROMPT_TEXT” }
}
}’

The call returned the following:

“candidates”: [
{
“content”: {
“role”: “model”,
“parts”: [
{
“text”: “Hello there! How can I help you today?”
}
]
},
“finishReason”: “STOP”,
“avgLogprobs”: -0.45475735664367678
}
],
“usageMetadata”: {
“promptTokenCount”: 1,
“candidatesTokenCount”: 10,
“totalTokenCount”: 33,
“trafficType”: “ON_DEMAND”,
“promptTokensDetails”: [
{
“modality”: “TEXT”,
“tokenCount”: 1
}
],
“candidatesTokensDetails”: [
{
“modality”: “TEXT”,
“tokenCount”: 10
}
],
“thoughtsTokenCount”: 22
},
“modelVersion”: “gemini-2.5-flash”,
“createTime”: “2026-07-29T10:31:34.164297Z”,
“responseId”: “htZpasmDCrnyxs0PubWNyAM”
}