Fix Azure OpenAI 429 insufficient_quota (2026 Guide)

The Azure OpenAI 429 insufficient_quota error confuses developers because it looks like a billing problem but usually isn’t. Azure has a two-layer limit system that most tutorials don’t explain clearly: your subscription’s credits (billing) and your deployment’s TPM (rate limit). This 2026 guide walks you through diagnosing which one you’re hitting and how to fix each in 5 minutes.

The exact error you’re seeing

openai.RateLimitError: Error code: 429
{
  "error": {
    "code": "insufficient_quota",
    "message": "You exceeded your current quota, please check your plan and billing details."
  }
}

This message is misleading. It almost never means “you ran out of money.” It usually means one of three things: your deployment’s TPM (tokens per minute) rate limit was hit, your Azure subscription has no OpenAI access yet, or your API key is pointing at OpenAI’s public API instead of Azure OpenAI.

Cause 1: You’re hitting TPM (tokens per minute) rate limit

Every Azure OpenAI deployment has a TPM quota. Default new deployments get 30k-120k TPM depending on model. If your app sends bursts of requests faster than the quota allows, you get 429.

Check your current TPM in Azure Portal:

1. Go to portal.azure.com.
2. Open your Azure OpenAI resource.
3. Click “Resource Management” > “Model deployments”.
4. Click on your specific deployment (e.g., “gpt-4o-mini-prod”).
5. The TPM limit shows at the top. Default is often 30k for gpt-4o-mini and 120k for gpt-4o.

To increase it:

1. In the same deployment view, click “Edit deployment”.
2. Under “Tokens per Minute Rate Limit (thousands)”, raise the slider to your subscription’s cap.
3. Click “Save and close”.

# Verify TPM via Azure CLI
az cognitiveservices account deployment show \
  --name your-openai-resource \
  --resource-group your-rg \
  --deployment-name gpt-4o-mini-prod \
  --query "sku"

Cause 2: Your subscription has no OpenAI access yet

Azure OpenAI requires approval for each subscription. If your subscription is new, you may not have OpenAI access even though the resource looks provisioned.

Check access:

1. In Azure Portal > your Azure OpenAI resource > Overview.
2. Look at the “Endpoint” URL. It should look like https://your-resource.openai.azure.com/.
3. Click “Model deployments”, if the list is empty, you have not deployed a model yet.
4. Click “Create new deployment”. If the model dropdown is empty or greyed out, your subscription lacks OpenAI access.

To request access: fill out Microsoft’s Azure OpenAI Limited Access application. Approval takes 1-3 business days for Enterprise Agreement subscriptions and 3-7 days for Pay-as-you-go.

Cause 3: Wrong endpoint or API key

The Azure OpenAI SDK uses different parameters than the public OpenAI SDK. If you’re calling openai.OpenAI with an OpenAI-style API key instead of openai.AzureOpenAI, you get 429 from OpenAI’s public API (not Azure).

Correct Azure OpenAI setup in Python:

from openai import AzureOpenAI
import os

client = AzureOpenAI(
    api_key=os.environ["AZURE_OPENAI_KEY"],
    api_version="2024-10-21",
    azure_endpoint="https://your-resource.openai.azure.com/"
)

response = client.chat.completions.create(
    model="gpt-4o-mini-prod",  # your deployment name, NOT the model name
    messages=[{"role": "user", "content": "Hello"}]
)

Common mistakes: (a) using openai.OpenAI instead of openai.AzureOpenAI, (b) passing the model name like “gpt-4o-mini” instead of your deployment name, (c) missing api_version.

Cause 4: You genuinely ran out of credits (rare)

If your Azure subscription is a free trial or pay-as-you-go without a payment method, you can exhaust credits. Check billing:

1. Azure Portal > Cost Management + Billing.
2. Look at “Cost analysis” for your subscription.
3. If the current spend is at or above your budget alert threshold, add a payment method or raise the budget.

For most Enterprise Agreement customers, this is not the cause. For personal accounts with USD 200 free credit, it can be.

Add retry logic to handle 429s gracefully

Even after raising TPM, bursty workloads will hit 429s occasionally. Handle them with exponential backoff:

import time
from openai import AzureOpenAI, RateLimitError

client = AzureOpenAI(api_key="...", api_version="2024-10-21", azure_endpoint="...")

def call_with_retry(messages, max_retries=5):
    for attempt in range(max_retries):
        try:
            return client.chat.completions.create(
                model="gpt-4o-mini-prod",
                messages=messages
            )
        except RateLimitError as e:
            wait = 2 ** attempt  # 1, 2, 4, 8, 16 seconds
            print(f"429 hit, retry {attempt+1} in {wait}s")
            time.sleep(wait)
    raise Exception("Max retries exceeded")

For production: consider Provisioned Throughput Units (PTU)

If your app consistently sends high volume, buy PTUs instead of relying on shared quota. PTUs give you dedicated capacity with predictable latency.

1. In Azure Portal > your Azure OpenAI resource > “Deployment types”.
2. Choose “Provisioned” instead of “Standard”.
3. Select model + PTU count (each PTU = ~50 TPS for GPT-4o).
4. Commit for 1 month or 1 year (annual commitment gets ~40% discount).

PTUs cost significantly more than pay-as-you-go but eliminate rate limit unpredictability. Suitable for production RAG systems, agent workflows, or customer-facing apps with SLA requirements.

Verification checklist

# 1. Confirm your deployment endpoint is Azure (not OpenAI public)
python -c "
from openai import AzureOpenAI
c = AzureOpenAI(api_key='...', api_version='2024-10-21',
                azure_endpoint='https://your-resource.openai.azure.com/')
print(c.chat.completions.create(model='your-deployment-name',
      messages=[{'role':'user','content':'test'}]).choices[0].message.content)
"

# 2. Check TPM usage vs quota via Azure Monitor
az monitor metrics list \
  --resource "/subscriptions/YOUR-SUB/resourceGroups/YOUR-RG/providers/Microsoft.CognitiveServices/accounts/YOUR-RESOURCE" \
  --metric "ProcessedPromptTokens,ProcessedCompletionTokens" \
  --interval PT1M