Azure OpenAI's 429s work differently from direct OpenAI or Anthropic, because quota here is assigned per deployment and pooled at the subscription level, not tied to your account as a whole. That difference changes how you diagnose and fix it.
Your quota is scoped to a subscription and a deployment, not your account
Tokens-per-minute and requests-per-minute limits are assigned per model deployment. A Global Standard deployment shares one quota pool across all regions in your subscription; a Data Zone Standard deployment shares one pool per data zone. A different deployment of the very same model has its own separate allowance entirely — so a 429 on one deployment says nothing about whether another has headroom.
You can hit a 429 even when your usage dashboard looks fine
Microsoft's own documentation notes you can see 429s even when token usage metrics appear below quota, because of how burst traffic is routed and handled — a sharp short spike, not your sustained average usage, is often the real trigger, and standard deployment types are built for low-to-medium, non-bursty traffic.
Fixing it
Three real levers: request a quota increase through the Azure quota request form; move quota between your own deployments if one is under-used and another is starved; or, for latency-critical or high-volume workloads, move to Provisioned Throughput Units (PTU) for dedicated, predictable capacity that sits outside the shared standard-tier quota pool entirely.