GPT-5.6 Luna on Foundry: PTU Sizing, PayGo vs. PTU + Spillover Pricing

GPT-5.6 Luna is a model available on Microsoft Foundry's platform, which offers different pricing models such as Provisioned Throughput (PTU) and PayGo (standard). Provisioned Throughput provides a fixed amount of processing capacity for a model deployment, with no sharing across tenants and a specific latency SLA. PayGo, on the other hand, charges based on actual usage.

To estimate the required PTUs for GPT-5.6 Luna, you need to consider factors like peak requests per minute (RPM), average input and output token sizes, cache rates, and model-specific constants. These constants include Input TPM per PTU, output-to-input ratio, and the minimum Global Provisioned deployment PTUs.

The process involves calculating Input TPM (input tokens per minute), Output TPM (output tokens per minute), Normalized TPM (normalized workload-specific tokens), and finally, the Estimated PTUs by dividing Normalized TPM by Input TPM per PTU. The PTUs deployed are rounded up to the nearest 5-PTU increment while ensuring they meet the minimum required PTUs.

When caching is involved, Effective Input TPM is reduced based on the cache rate, resulting in lower Estimated PTUs. This cache optimization can significantly reduce costs. The sample calculations demonstrate the impact of a 50% cache hit rate, reducing the required PTUs from 80 to 60.

The article also provides an illustrative example of how traffic patterns vary throughout the day, with peak hours and a 30-day estimate. PayGo pricing calculates costs based on Azure OpenAI's USD rates for gpt-5.6-luna, factoring in regular input, cached input, cache writes, and output token prices. Comparing these two pricing models helps users make informed decisions based on their workload requirements and cost considerations.