Azure OpenAI Service Commitment: When PTU Pays and When It Does Not

At June 2026 list, a Global PTU annual reservation undercuts Pay As You Go only above roughly 85 percent sustained utilization on GPT 5. Most enterprise estates we benchmark sit below 50 percent.

Prepared by Redress Compliance  ·  June 2026  ·  Representative Azure OpenAI estate scenario (benchmark scenario, not a quote)

Azure OpenAI is rarely bought alone. It sits inside the Azure consumption commitment (MACC) and often inside an Enterprise Agreement that also carries Microsoft 365 and Copilot. That placement is deliberate: both Pay As You Go tokens and PTU reservations decrement the MACC, so every Azure OpenAI dollar accelerates your committed cloud burn.

This is why the account team pushes PTU conversion even when the workload math does not support it. A reservation pulls twelve months of AI spend into the current MACC period, helps the commitment retire on schedule, and anchors the next, larger MACC renewal.

The buyer side move is separation. Negotiate the Azure OpenAI commitment as its own conversation with its own utilization evidence, then decide how it nets against the MACC. Customers who let the two blend accept commitment sizes the standalone numbers would have rejected.

Azure OpenAI prices through two mechanics. Pay As You Go bills per million input and output tokens, with the catalog spanning the GPT 5 family, GPT 4.1, the o series reasoning models, and embedding, image, and audio models. A Batch tier prices asynchronous workloads at roughly half the interactive rate, and cached input tokens bill at a 90 percent discount.

Provisioned Throughput Units reserve model processing capacity at a fixed hourly rate per PTU, regardless of usage. PTUs are billed hourly, with one month and one year Azure reservations providing the term discount. Three deployment types carry different rates and data boundaries.

Throughput per PTU is set per model. GPT 5 delivers 4,750 input tokens per minute per PTU, and one output token counts as eight input tokens against utilization, matching the 8 to 1 price ratio. Cached tokens are deducted 100 percent from PTU utilization, so aggressive prompt caching effectively expands reserved capacity for free.

Read that table twice: PTU billing is model independent, but PTU capacity is not. The same reservation that runs GPT 5 at 4,750 tokens per minute per PTU runs GPT 5.5 at 1,200, so a model upgrade can cut delivered throughput by three quarters at the same price. Section 4 turns that mechanic into contract language.

Start from one anchor: at GPT 5 list of $1.25 per million input tokens and $10 per million output tokens, Pay As You Go costs $1.25 per million utilization weighted tokens regardless of mix. One fully utilized PTU processes about 208 million such tokens per month. So $260 of Pay As You Go buys a full PTU month.

Against that, the three PTU billing modes price as follows.

The monthly reservation at $260 exactly equals the Pay As You Go value of a fully utilized PTU, so it breaks even only at 100 percent utilization. The annual reservation at $221 breaks even at roughly 85 percent. Below those lines Pay As You Go is cheaper; above them the reservation wins and adds the latency floor.

A worked estate, priced four ways

Take a representative insurance carrier running GPT 5 document workloads: 1,200 million input tokens and 150 million output tokens per month, peaking at twice the average rate. Utilization weighted demand is 2,400 million tokens. Sizing for peak requires 25 Global PTUs, which run at 46 percent average utilization (benchmark scenario, not a quote).

The right structure for most estates is a commitment ladder: an annual reservation sized to the measured utilization floor (the demand present in 85 percent or more of hours), a monthly reservation tranche for seasonal steps, and Pay As You Go spillover above the reservation, which Azure bills automatically at the hourly rate when deployments exceed reserved PTUs.

Every Azure OpenAI commitment is a bet against the price curve. The flagship input token list has fallen relentlessly: GPT 4 launched at $30 per million input tokens in March 2023, GPT 4 Turbo cut that to $10, GPT 4o to $5, the August 2024 GPT 4o update to $2.50, and GPT 5 to $1.25 in August 2025.

Three clauses convert that curve from risk into protection. First, a model price ceiling: negotiated token rates expressed as a discount off the then current public list, so every public price cut flows through automatically. A fixed dollar rate in a falling market is a concession to Microsoft.

Second, model substitution rights with throughput protection. Because tokens per PTU vary by model, a substitution clause must hold the delivered tokens per minute, not just the PTU count. Moving a 25 PTU GPT 5 deployment to GPT 5.5 without that language cuts delivered throughput by roughly three quarters at identical cost.

Third, deprecation and retirement notice. Microsoft's published lifecycle policy guarantees generally available models for at least 12 months and commits to at least 60 days notice before retirement. For a regulated workload, negotiate 90 to 180 days in the agreement, plus funded migration assistance when retirement forces a model move inside the reservation term.

The least understood sentence in the PTU documentation is this one: quota does not guarantee capacity, and neither does a reservation. PTU quota caps what you may deploy in a region; the reservation guarantees a discounted price. Neither guarantees that GPT 5 capacity exists in your region on the day you scale.

Microsoft's own guidance is to deploy first, then buy the reservation, precisely because customers have purchased reservations against capacity that was not available. Cancellation credits on reservations are limited, and reservations for Global, Data Zone, and Regional deployment types are not interchangeable with each other.

The buyer side procedure is a regional capacity audit before any commitment signs:

Then negotiate capacity language: a documented capacity escalation path with named engineering contacts, priority access for committed customers during regional constraint, and the right to rescope the reservation to another region without penalty if Microsoft cannot provision within 30 days.

The strongest commercial argument for Azure OpenAI over a direct OpenAI contract is the Customer Copyright Commitment in the Microsoft Product Terms. Microsoft commits to defend customers against third party copyright claims over output content and to pay resulting judgments or settlements.

The protection is conditional, and the conditions are where coverage quietly dies. To be covered for Azure OpenAI output, the customer must have implemented all required mitigations documented by Microsoft, including the content filters and safety system configuration. An engineering team that relaxes content filters for throughput can void the indemnity for the whole workload.

The carve outs matter as much. The commitment does not cover claims arising from your input data, your fine tuning training sets, or material you supply through retrieval augmented generation. Your data pipeline remains your liability.

Estates that run both Microsoft 365 Copilot and Azure OpenAI face a deliberate blending play. Copilot lists at $30 per user per month, $360 per user per year. The account team will present Copilot seats and the Azure OpenAI commitment as one AI envelope, sized together, discounted together.

Keep them apart, because the economics are different in kind. Copilot is a per seat product whose value depends on adoption per user. Azure OpenAI is metered infrastructure whose value depends on workload engineering.

That is not an argument against Copilot, which bundles the Microsoft 365 graph integration no custom build matches cheaply. It is an argument for routing each use case to its cheapest adequate vehicle: broad knowledge worker assistance to Copilot seats under an adoption gated ramp, custom and high volume workloads to Azure OpenAI under the commitment ladder.

These are the clauses we negotiate into Azure OpenAI commitments, in the order they typically trade.

Sequence matters as much as substance. The levers price differently at different points in Microsoft's quarter, and utilization evidence is the currency that buys them.