Microsoft Foundry in Practice

Requesting a Quota Increase for Claude on Foundry

Default rate limits on pay-as-you-go are sized for pilots, not production. When you outgrow them, there's one official channel — and a little preparation makes the request go smoother.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Claude rate limits on Microsoft Foundry are set per model at the Azure subscription level, measured in Requests Per Minute (RPM) and uncached Input Tokens Per Minute (ITPM). The defaults depend on your agreement type, and the gap is large: a pay-as-you-go subscription gets 40 RPM / 40,000 ITPM for Opus-family models and Sonnet 5 (80 RPM / 80,000 ITPM for Sonnet 4.6, Sonnet 4.5, and Haiku 4.5), while enterprise agreements (MCA-E) start at 2,000 RPM / 2,000,000 ITPM for those same top-tier models and 4,000 / 4,000,000 for the smaller ones. If your defaults can't carry your projected traffic, this article covers the official way up.

First, confirm you actually need more quota

Before filing anything, rule out cheaper fixes. Check the Foundry portal's Monitoring tab for per-model token and request consumption and identify which dimension you're hitting — RPM or ITPM (they throttle independently; see Foundry quota types). If it's ITPM, prompt caching can change the math substantially: ITPM counts uncached input tokens and cache writes, but cache reads are free against quota, so a cached system prompt stops counting after the first request. Also remember quota pools are shared across all deployments of the same model and version in a subscription — if dev and prod share a subscription, separating them may relieve production without any increase.

The official channel

Quota increases beyond the defaults go through Microsoft's quota increase request form, linked from the Claude models documentation at aka.ms/oai/stuquotarequest. This is a Microsoft process, not an Anthropic one — Claude on Foundry is an Azure Marketplace offering, and Microsoft administers the subscription-level limits.

Two situations make the request effectively mandatory rather than optional:

What to have ready

Microsoft's public documentation doesn't enumerate the form's required fields, so treat the following as sensible preparation rather than an official checklist. Requests that map cleanly to real usage are easier to evaluate, so bring:

Rule of thumb: ask for what a realistic six-month projection needs, not a 100x moonshot. Grounded requests backed by Monitoring-tab data are easier to approve than round numbers pulled from the air.

Timelines and planning

Microsoft does not publish a service-level turnaround for Claude quota requests, so build the request into your project plan rather than your launch week — file as soon as a pilot's numbers make production traffic estimable. If your organization is negotiating an enterprise agreement anyway, note that MCA-E limits are 50x the pay-as-you-go defaults out of the box, which may make an interim request unnecessary. While a request is pending (and after it's granted), keep exponential backoff on 429s in place: Foundry doesn't return Anthropic's anthropic-ratelimit-* headers, so graceful throttling behavior is your application's job either way — see rate limit handling.

Where to go next

Understand the counting rules in Foundry quota types, squeeze more from existing quota with prompt caching, and compare the equivalent processes on other clouds in quotas and limits across platforms.

Sources