Within the last 24 hours, an identical, previously working multimodal .jsonl pipeline has started failing consistently when submitted via Vertex AI Batch Prediction (projects.locations.batchPredictionJobs) with model gemini-3-flash-preview. The exact same prompt and source files succeed flawlessly when executed via the Online/On-Demand endpoint (projects.locations.endpoints.predict) with model gemini-3-flash-preview.
The exact task succeeds in batch prediction and online endpoint when switching to gemini-3.1-flash-lite.
Error in batch line response in batch output file:
“status”:“{“code”:3,“space”:“generic”,“message”:“could not find @type ‘type.googleapis.com/beyond.IntrospectionClassifierResult’”}”
Batch error:
Code: 13, Message: System error. Please try this operation again. If the issue persists please visit Support Hub | Google Cloud to view your support options.
@Google- Please check this bug.