Capacity planning for Claude on Microsoft Foundry has an unusual step zero: check what kind of Azure subscription the workload will bill through. Microsoft sets default Claude rate limits per subscription type, and the difference between a pay-as-you-go subscription and an Enterprise agreement is not a modest bump — it is roughly fifty-fold on the models where both are available, and infinite on Claude Fable 5, which pay-as-you-go subscriptions cannot use at all by default.
The numbers, side by side
Foundry measures Claude rate limits in requests per minute (RPM) and uncached input tokens per minute (ITPM). The defaults below are from Microsoft's Claude models documentation as of July 2026 and are subject to change; always check the live page before committing to a capacity plan.
| Model group | Pay-as-you-go default | Enterprise / MCA-E default |
|---|---|---|
| Opus-family and Sonnet 5 | 40 RPM · 40,000 ITPM | 2,000 RPM · 2,000,000 ITPM |
| Sonnet 4.6, Sonnet 4.5, Haiku 4.5 | 80 RPM · 80,000 ITPM | 4,000 RPM · 4,000,000 ITPM |
| Claude Fable 5 | 0 RPM · 0 ITPM | 2,000 RPM · 2,000,000 ITPM |
| Free Trial | 0 across the board | |
Read the pay-as-you-go column carefully. Forty requests per minute of Opus 4.8 is a pilot, not a production allocation — one modestly busy internal tool can saturate it. And the Fable 5 row is a hard zero: a team on a pay-as-you-go subscription that architects around Fable 5 will discover at deployment time that its default quota does not permit a single request (see Foundry and Fable 5's zero default quota). That is why subscription type belongs in the platform-selection checklist, not the launch checklist.
How Foundry counts your ITPM
The ITPM figure has Foundry-specific accounting: it counts uncached input tokens plus 5-minute and 1-hour cache-write tokens. Cache reads do not count, and output tokens (including cache reads) do not count either. Practically, that means a cache-heavy workload — a long shared system prompt with small per-request suffixes — consumes far less of its ITPM allowance than raw traffic volume suggests, which is one of the few levers available for stretching a pay-as-you-go allocation while a quota increase or contract change is in flight.
Also remember the pool is shared: quota is managed at the Azure subscription level, and all Global Standard deployments of the same model and version in a subscription draw from one pool across regions. Spinning up more deployments does not create more quota (details here).
What to do about it
Confirm the subscription type first. Claude on Foundry already requires a paid Azure subscription — free trial, student, and credit-only subscriptions are not supported for deployment at all. Within paid subscriptions, the rate-limit table above then splits pay-as-you-go billing from Enterprise and MCA-E (Microsoft Customer Agreement — Enterprise) contracts. If your organization has an Enterprise agreement, make sure the Claude workload actually bills through it rather than through a team's ad hoc pay-as-you-go subscription.
Request increases through Microsoft's form. Quota increases beyond the defaults go through Microsoft's quota increase request form (linked from the Claude models documentation), not through Anthropic. Plan lead time for this — see requesting a quota increase on Foundry.
Instrument for throttling, not headers. Foundry does not return Anthropic's anthropic-ratelimit-* response headers, so your application cannot watch remaining capacity per request the way it can on the Claude API. Monitor 429 rates and token metrics in Azure Monitor and back off exponentially on throttle responses (how to monitor quota without the headers).
Right-size the model. The 2x default headroom for Sonnet 4.6/4.5 and Haiku 4.5 over the Opus family exists at both subscription levels. If part of the workload passes quality evaluation on a smaller model, moving it doubles that traffic's ceiling — the same logic as right-sizing model choice to reach a throughput target.
Where to go next
Committed throughput vs pay-as-you-go on Foundry covers the contractual side, and the cross-platform quota guide puts Foundry's model in context against Bedrock and Vertex.