Azure AI Foundry Rate Limits

Note

Updated September 2026. Microsoft has rebranded Azure AI Foundry to Microsoft Foundry; Claude models went GA (~July 2026) with broader subscription eligibility and a cache-aware rate limit model. This page supersedes earlier versions that described Claude as EA/MCA-E-only with zero pay-as-you-go quota. With the gateway-centric architecture, FoundryGate now actively mitigates these limits rather than passing 429s through — but the platform ceilings below still determine total scale.

The two-layer picture

• FoundryGate quotas govern who gets access and how much they consume per month — enforced in real time at the APIM gateway (llm-token-limit).
• Foundry rate limits govern how fast the platform processes requests right now, shared across callers. This page is about the second layer.

The gateway smooths and reroutes around transient limit hits (per-developer TPM caps, backend pools, circuit breakers honoring Retry-After), but it cannot manufacture capacity beyond what the limits below allow.

---

Claude models on Microsoft Foundry

Eligibility

Claude requires a paid Azure subscription with an active pay-as-you-go billing method in an Anthropic-supported billing country. Not supported: CSP subscriptions, free trial / student / startup-credit subscriptions, credit-only sponsored subscriptions, and EA accounts in South Korea. This is a real broadening from the earlier EA/MCA-E-only rule — but check your billing method before planning a rollout. The deploying principal also needs Azure Marketplace permissions (Microsoft.SaaS/register/action and related) for the automated offer attestation.

How Claude limits are measured (cache-aware)

Claude limits are expressed as three numbers per model, pooled per subscription per model (shared across all Global Standard deployments of that model in every region):

• RPM — requests per minute
• ITPM — uncached input tokens per minute (cache reads are free against ITPM)
• OTPM — output tokens per minute

Tip

The cache-aware accounting matters enormously for Claude Code, which caches system prompts and conversation prefixes aggressively. Effective input capacity is far higher than raw ITPM suggests — often the binding constraint becomes OTPM or RPM, not input tokens.

Default limits (September 2026)

Model	PAYG (RPM / ITPM / OTPM)	Enterprise & MCA-E (RPM / ITPM / OTPM)
claude-opus-5, claude-opus-4-x	40 / 40K / 8K	2,000 / 2M / 400K
claude-sonnet-5	40 / 40K / 8K	2,000 / 2M / 400K
claude-sonnet-4-6, claude-sonnet-4-5	80 / 80K / 16K	4,000 / 4M / 800K
claude-haiku-4-5	80 / 80K / 16K	4,000 / 4M / 800K

Free-trial subscriptions get zeros. Quota increases go through the Azure OpenAI/Foundry quota request form — Claude is the only partner model family eligible for increases, they are form-based (not programmatic), and priority goes to subscriptions actively consuming existing quota.

Deployment types, hosting, and regions

• Deployment types: Global Standard and Data Zone Standard (US only). No PTU / provisioned option exists for Claude — and therefore no spillover and no priority processing (both are Azure OpenAI-only features).
• Hosting variants are encoded in the model version: version 1 = hosted on Anthropic infrastructure (eastus2, swedencentral), version 2 = hosted on Azure (GA; available across roughly nine US/EU regions, plus Data Zone Standard US).
• A possible 300-concurrent-request cap: the generic Foundry Models limits table still lists 300 concurrent requests for partner models, but Claude’s own documentation now expresses limits only as RPM/ITPM/OTPM. Treat the cap as unverified and test at scale before assuming either way.
• Claude on Foundry does not return anthropic-ratelimit-* response headers. FoundryGate surfaces the gateway’s own remaining-quota headers instead.

What Enterprise defaults support in practice

A worked example for claude-sonnet-5 at Enterprise defaults (2,000 RPM / 2M ITPM / 400K OTPM), assuming heavy agentic coding (6 requests/min per active developer, ~800 output tokens/request, well-cached prompts averaging ~1.5K uncached input tokens/request):

Constraint	Capacity	Requests/min it allows
RPM	2,000	2,000
ITPM (uncached)	2,000,000	~1,330
OTPM	400,000	~500

OTPM binds first: ~500 requests/min ≈ 80–85 simultaneously heavy developers, or a total org of several hundred developers at typical concurrency ratios (15–30% active). Poorly cached workloads shift the binding constraint to ITPM and cut this roughly in half. If the 300-concurrent cap applies, long-running streamed requests can bind before OTPM does.

---

Azure OpenAI models (GPT / Codex families)

GPT-family models use the tiered quota system (automatic tier progression by subscription type and usage) and support the full traffic-management toolbox that Claude lacks: PTU (provisioned throughput), spillover from provisioned to standard (GA), priority processing, dynamic quota (preview), and Global Batch.

For the Codex CLI models specifically (deployable from the Foundry catalog: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.2-codex, gpt-5.3-codex):

• Reported Global Standard defaults range from ~15K TPM (gpt-5.1-codex) to ~150K TPM (gpt-5.3-codex, gpt-5.1-codex-max); gpt-5.1 chat models default around 300K TPM (1M enterprise).
• These defaults are volatile — Microsoft has cut them without notice (community reports of 300K→50K on gpt-5.1). Check az cognitiveservices usage list -l <region> against your subscription rather than trusting any published table, including this one.

Note

Older GPT-4.1/GPT-4o tier tables that previously appeared on this page predate the gpt-5 lineup and have been removed; the tier mechanics are unchanged, the specific numbers were stale.

---

How FoundryGate mitigates transient limits

The gateway architecture addresses the symptoms in order:

1. Per-developer TPM caps (llm-token-limit, per subscription key) — one developer’s runaway agentic loop gets a targeted 429 + Retry-After (both Claude Code and Codex CLI back off automatically) instead of exhausting the shared pool.
2. Backend pools + circuit breakers — when a deployment itself 429s, its breaker trips (honoring Retry-After) and requests retry against the next deployment in the pool (other region, Data Zone, or hosting variant). Clients see latency, not errors.
3. Deployment resizing — deployment sku.capacity (thousands of TPM) is PATCH-able at runtime; admins rebalance capacity between deployments from FoundryGate without redeploying.
4. Right-sizing from data — the ApiManagementGatewayLlmLog reconciliation feed shows whether you are RPM-, ITPM-, OTPM-, or concurrency-bound; each has a different fix (quota increase, caching discipline, output caps, more pool members).

What no gateway can do: exceed subscription-level pools, or buy dedicated Claude throughput (PTU does not exist for Claude). Claude-heavy organizations beyond a few hundred active developers should plan on quota-increase requests early — or a direct Anthropic API agreement for the heaviest cohort.

Planning guide

Team size	Approach
< 50 developers	PAYG limits are workable for pilots on sonnet-4-6/haiku (80 RPM); Enterprise defaults comfortable. Single deployment per model, gateway TPM caps on.
50–200	Enterprise/MCA-E defaults + multi-deployment pools (Global Standard + Data Zone). File a Claude quota increase before you need it. GPT/Codex: watch volatile defaults, consider PTU for the busiest model.
200–500	Claude: quota increases mandatory; OTPM is your likely binding constraint — enforce output caps at the gateway. GPT: PTU + spillover.
500+	Claude via Foundry becomes a structural constraint; split the heaviest users to a direct Anthropic agreement, keep the long tail on Foundry. GPT: PTU per region behind pools.