Calling Gemini: Auth, SDK, Quotas, Cost and Safety

GCP Basics · AI on GCP – Gemini, RAG, BigQuery ML and Agents · Lesson 2 of 5

Call Gemini with the Google Gen AI SDK or REST using Application Default Credentials and the Vertex AI User role; plan for tokens, quotas and 429s, and use safety settings and Model Armor to filter harmful input and output.

Calling Gemini on Google Cloud is an authenticated API call from your code. No API key in the code: use Application Default Credentials and IAM.

flowchart LR
  APP["Cloud Run service<br/>runs as chat-sa"] -- "ADC token" --> API["aiplatform.googleapis.com<br/>location global"]
  API --> MA["Model Armor<br/>prompt injection, sensitive data"]
  MA --> GEM["Gemini model"]
  GEM --> SAFE["Safety filters"]
  SAFE --> APP
  %% step APP: The service runs as chat-sa, which has the Vertex AI User role; the SDK gets tokens automatically from the metadata server.
  %% step API: Requests go to the Agent Platform / Vertex AI API, here on the global endpoint.
  %% step MA: Optionally, Model Armor screens prompts for prompt injection, jailbreaks and sensitive data before they reach the model.
  %% step GEM: The model generates a response; you pay for input and output tokens.
  %% step SAFE: Built-in safety filters can block harmful content in the response by category and threshold.

Authentication and roles

Grant the calling identity Vertex AI User (roles/aiplatform.user) on the project.
On Cloud Run, GKE or a VM, the attached service account is used automatically; locally, gcloud auth application-default login.
Avoid API keys in code; if a key is unavoidable for a prototype, restrict it and keep it in Secret Manager.

SDK and REST

# pip install google-genai
# env: GOOGLE_GENAI_USE_VERTEXAI=True GOOGLE_CLOUD_PROJECT=dl-ai-dev GOOGLE_CLOUD_LOCATION=global
from google import genai
client = genai.Client()
resp = client.models.generate_content(
    model="gemini-3.5-flash",   # check the current model list for IDs
    contents="Summarise this customer complaint in one sentence: …")
print(resp.text)

curl -X POST -H "Authorization: Bearer $(gcloud auth print-access-token)" -H "Content-Type: application/json"   https://aiplatform.googleapis.com/v1/projects/dl-ai-dev/locations/global/publishers/google/models/gemini-3.5-flash:generateContent   -d '{"contents":[{"role":"user","parts":[{"text":"Hello"}]}]}'

Tokens, cost and quotas

Topic	What to know
Pricing	Per million input and output tokens; Flash-Lite < Flash < Pro
Context caching	Reuse a long prompt prefix at a lower price
Batch prediction	Large offline jobs at about half the online price
Quotas	Many Gemini models use dynamic shared quota: no fixed per-project limit, but HTTP 429 when capacity is busy; retry with exponential backoff
Provisioned Throughput	Reserve capacity for predictable latency and volume

Set max output tokens, keep prompts short, and log token counts (usage_metadata) to watch cost.

Safety

Safety settings per request: harm categories (harassment, hate speech, sexually explicit, dangerous content) with block thresholds.
Model Armor screens prompts and responses for prompt injection, jailbreak attempts, sensitive data (via Sensitive Data Protection) and malicious URLs, across models.
Use system instructions to set the assistant’s role and limits; never trust user input to override them.

Command	Meaning
`gcloud projects add-iam-policy-binding dl-ai-dev --member=serviceAccount:chat-sa@dl-ai-dev.iam.gserviceaccount.com --role=roles/aiplatform.user`	Allow the app to call models
`gcloud auth print-access-token`	Token for a quick curl test

Traps: a 429 is usually capacity, not a bug, so add backoff; long chat histories are re-sent and billed as input tokens on every turn; outputs can be wrong or unsafe, so validate them before acting on them.