Google Gemini
Rate Limits
HTTP 503
Gemini 503 model_overloaded
503 model_overloaded
Gemini's servers are overloaded and can't serve your request right now. This isn't your quota — it's Google's capacity. Retry with backoff or fall over to Flash or a different provider.
What you see when this happens
- HTTP 503 with 'model is overloaded' or 'try again later'
- Correlates with Google Cloud status page reports
- Different models affected differently
- May be regional (one Vertex region affected)
Common causes
Global Gemini capacity saturation
Peak usage periods hit capacity limits.
Specific model overloaded
Pro or Ultra may be overloaded while Flash still serves.
Regional issue
Some GCP regions may hit capacity before others.
How to fix it
Retry with exponential backoff
Standard retry pattern.
from google.genai import errors
import random, time
for attempt in range(5):
try:
return client.models.generate_content(
model="gemini-3.5-pro", contents=prompt,
)
except errors.APIError as e:
if e.code == 503 and attempt < 4:
time.sleep(random.uniform(0, min(60, 2 ** attempt)))
continue
raise
Fall over to Gemini Flash
Flash has separate capacity from Pro — often available when Pro is overloaded.
Fall over to a different provider
Multi-provider fallback (Gemini → Claude → GPT) preserves uptime.
Try a different Vertex AI region
If one region is overloaded, another may still serve.
How to prevent it
- Monitor status.cloud.google.com for Vertex AI incidents
- Build multi-provider fallback
- Use cross-region deployments for Vertex AI production workloads
Related errors
Deeper reading
Frequently asked questions
How does 503 differ from 429 on Gemini?
429 = your quota exhausted (fix: wait or upgrade). 503 = Google's capacity exhausted (fix: wait or fail over). Different root causes, different fixes.