Microsoft Foundry in Practice

Committed Throughput vs. Pay-As-You-Go on Foundry

Teams arriving from Azure OpenAI or Bedrock often ask where the provisioned throughput SKU is for Claude. The honest answer: as of July 2026, Microsoft's docs don't describe one — commitment on Foundry works through different levers.

Claude 3P 101 · Updated July 2026 · Unofficial guide

"Committed throughput" usually means paying up front for reserved model capacity — a fixed hourly fee for guaranteed requests-per-minute, as with Bedrock's provisioned throughput or Azure OpenAI's provisioned units. For Claude on Microsoft Foundry, the documented billing model points the other way: usage bills in Claude Consumption Units (CCU) through Azure Marketplace, metered hourly and invoiced monthly in arrears, explicitly with no prepaid CCU credits or balances. Microsoft's Claude documentation does not describe a reserved-capacity or provisioned-throughput SKU for Claude deployments. If your capacity plan assumes one exists, verify with the official documentation and your Microsoft account team before building around it.

That does not mean "pay-as-you-go or nothing." Foundry has three real commitment levers — they just address capacity, price, and procurement separately rather than bundling them into one SKU.

Lever 1: agreement type sets your throughput ceiling

Rate limits on Foundry are quota, not a purchase. They are measured in requests per minute (RPM) and uncached input tokens per minute (ITPM), managed at the subscription level and shared across deployments of the same model and version. The defaults differ dramatically by agreement:

Model groupPay-as-you-go defaultEnterprise / MCA-E default
Opus family, Sonnet 5, Fable 540 RPM / 40,000 ITPM (Fable 5: 0 / 0)2,000 RPM / 2,000,000 ITPM
Sonnet 4.6, Sonnet 4.5, Haiku 4.580 RPM / 80,000 ITPM4,000 RPM / 4,000,000 ITPM

Note that Claude Fable 5 defaults to zero throughput on plain pay-as-you-go — enterprise terms or a quota grant are effectively required to use it at all. Beyond the defaults, increases go through Microsoft's quota increase request form. The quota costs nothing; you still pay only for tokens used.

Lever 2: private offers set your price

Volume discounts on Foundry are delivered as Azure Marketplace private offers, applied at the token-to-CCU conversion step, potentially with different rates per model. This is where the classic committed-use trade lives: in exchange for a commitment your organization negotiates, you accrue fewer CCUs per token. The break-even logic is the familiar one, applied to a negotiation rather than a SKU purchase:

Rule of thumb: negotiate a private offer once your organic monthly spend is steady, not before. Because Foundry billing is postpaid, running pay-as-you-go for a quarter to gather real numbers costs you nothing but list price — and produces the evidence you need at the negotiating table.

Lever 3: MACC absorbs the spend

The CCU meter is MACC-eligible: Claude spend decrements a Microsoft Azure Consumption Commitment like other qualifying marketplace consumption. For enterprises that have already committed cloud dollars, this can make Foundry pay-as-you-go behave financially like committed spend — the commitment exists at the agreement level, and Claude simply draws it down. Details are in how Azure Marketplace billing works for Foundry.

Decision summary

Choose plain pay-as-you-go when traffic is new, spiky, or experimental. Move to enterprise-agreement quotas when default RPM/ITPM limits throttle production (see requesting a quota increase). Pursue a private offer when monthly spend is predictable enough to commit against. And if your workload genuinely requires reserved, guaranteed capacity of the kind Bedrock sells, compare platforms before assuming — capabilities differ by cloud, and Microsoft's Claude docs currently document quota, not reservations.

Sources