GPT-5.6 Luna Pricing on Azure: PTU vs. PayGo and Spillover

Understanding Provisioned Throughput Units (PTUs) for GPT-5.6 Luna

Azure OpenAI's introduction of Provisioned Throughput Units (PTUs) offers a distinct approach to model deployment, particularly for demanding workloads. Unlike the standard Pay-As-You-Go (PayGo) model, PTUs provide a dedicated, fixed amount of processing capacity exclusively for your model deployment. This dedicated capacity comes with a model-specific latency Service Level Agreement (SLA), ensuring consistent performance that isn't shared across multiple tenants. This is crucial for applications where predictable response times are paramount, such as real-time conversational AI, complex data analysis pipelines, or critical business process automation.

The PTU model is ideal for predictable, sustained traffic with consistent latency and high-throughput requirements. If your application experiences consistent demand, the overhead of managing variable PayGo costs and the uncertainty of shared resource performance can be significant. PTUs eliminate this by reserving capacity, allowing developers and operations teams to engineer for a known performance envelope.

A key characteristic of PTUs is their model-agnostic nature regarding quota allocation. The same quota pool can be allocated across different supported models within a subscription and region. However, the actual throughput achieved per PTU remains specific to the model and its version. This means that while you manage a single pool of PTUs, a PTU allocated to GPT-5.6 Luna will provide a different throughput than the same PTU allocated to, say, GPT-4 Turbo. This flexibility allows for dynamic reallocation of resources as application needs evolve, provided the underlying models are compatible with the PTU framework.

PTU quota is granted on a granular level: per subscription, per region, and per deployment type. This means that quota allocated in one region, such as East US, does not carry over to another region like West Europe. Similarly, Global Provisioned quota does not automatically extend to Data Zone Provisioned capacity. Careful planning is required to ensure adequate PTU allocation across all necessary regions and deployment environments to avoid performance bottlenecks or unexpected costs.

PayGo vs. PTU: A Pricing and Sizing Comparison

The choice between PayGo and PTU pricing hinges on traffic patterns and performance demands. PayGo is the default and most flexible option for variable workloads. You are billed based on actual token consumption (input and negative output tokens), making it cost-effective for sporadic or low-volume usage. However, PayGo deployments share underlying infrastructure, meaning latency can fluctuate based on overall demand on Azure's OpenAI service. This makes it less suitable for applications requiring strict latency guarantees or consistent high throughput.

PTU pricing, conversely, involves a commitment to a fixed capacity over a period, typically billed hourly. This commitment guarantees a specific level of throughput and latency. While it may appear more expensive upfront for low-traffic scenarios, it can become more economical for high, consistent traffic. The cost per token effectively decreases with higher sustained usage compared to PayGo, due to the guaranteed capacity and avoidance of potential throttling or rate limiting inherent in shared environments.

Sizing PTUs requires careful analysis of expected peak and average throughput. Azure OpenAI provides guidance, but ultimately, it involves understanding your application's token processing needs per second. For GPT-5.6 Luna, specific throughput values per PTU will be published by Azure. A common approach is to monitor PayGo usage over a representative period, identify peak sustained token-per-second requirements, and then select a PTU allocation that comfortably exceeds this peak, accounting for potential future growth and the SLA requirements.

Spillover Pricing: Bridging the Gap

The concept of spillover pricing is critical for optimizing costs when using PTUs. Spillover allows a PTU deployment to utilize additional capacity beyond its provisioned quota, but at a different rate. When your provisioned PTU capacity is exhausted, requests can 'spill over' into a shared pool, typically billed at the standard PayGo rate. This mechanism provides a safety net, ensuring that your application can handle occasional spikes in demand without outright failure or severe throttling, while still benefiting from the predictable performance of your dedicated PTUs for the bulk of your traffic.

The benefit of spillover is that it prevents the need to over-provision PTUs to handle infrequent peak loads. Instead, you can provision for your typical high-demand scenario and allow the overflow to be handled by the more flexible, albeit potentially more expensive per-token, PayGo pool. This hybrid approach offers a balance between guaranteed performance and cost efficiency. The exact spillover rate and conditions are defined by Azure's specific terms for PTU deployments.