Azure OpenAI Pricing Calculator

Azure OpenAI charges appear on your regular Azure invoice alongside compute, storage, and other services. There is no separate bill from OpenAI. Charges are metered like any other Azure consumption resource, which simplifies accounting for organizations already managing Azure spending through cost centers, departments, or resource groups.

Azure OpenAI offers two billing models: pay-as-you-go and Provisioned Throughput Units (PTUs). Pay-as-you-go is the default — you pay per token with separate rates for input and output, no minimum commitments, and no upfront costs. This works well for development, testing, and variable workloads where request volume is unpredictable. The calculator above uses pay-as-you-go rates.

PTUs are a reserved capacity model for production workloads with sustained traffic. You purchase a fixed amount of throughput measured in PTUs, billed hourly regardless of utilization — similar to reserved VM instances. The trade-off: predictable costs and guaranteed throughput in exchange for paying for capacity you might not fully use. PTUs can reduce per-token costs by 50% or more at high utilization, but generally only make financial sense once you consistently use 60-70% or more of the provisioned capacity.

Azure OpenAI also distinguishes between Standard and Provisioned deployments, and between Global and Regional routing. Standard deployments use pay-as-you-go billing with shared infrastructure — best-effort throughput that may experience rate limiting under high demand. Provisioned deployments reserve dedicated compute with consistent latency, available with monthly or annual commitments. Global deployments route to the nearest data center for better availability; Regional deployments pin your model to a specific Azure region for data residency control. Per-token rates are the same for both Global and Regional Standard deployments.

The most common question when budgeting for Azure OpenAI is: how many tokens will my application actually use? Token counts depend on the use case, prompt length, and how much output you request. Below are practical ranges based on typical production workloads that you can plug into the calculator above.

Chat applications typically consume 500 to 2,000 input tokens per turn, including the system prompt, conversation history, and the user's message. As conversations grow longer, input tokens increase because the full history is sent with each request. A 10-turn conversation can reach 4,000 to 6,000 input tokens on the final turn. Output tokens per turn usually range from 200 to 1,000. To control costs, consider truncating or summarizing conversation history after a certain number of turns.

Document summarization scales directly with document length — a short document might use 1,000 to 2,000 input tokens, while a long report can consume 5,000 to 50,000 or more. Output tokens for summaries are modest, typically 200 to 800. Documents exceeding a model's context window require a chunking strategy, meaning multiple API calls and proportionally higher costs.

RAG (retrieval-augmented generation) adds retrieved context chunks to each prompt, pushing input tokens higher. A typical RAG request includes the system prompt, user query, and 3 to 10 retrieved text chunks, totaling 2,000 to 8,000 input tokens. This can reach 20,000 to 50,000 tokens with large passages or many chunks.

Code generation typically uses 500 to 3,000 input tokens and 300 to 2,000 output tokens. Complex tasks with extensive codebase context push input tokens much higher. Classification and extraction tasks are the most token-efficient, often using only 200 to 1,000 input tokens and 50 to 200 output tokens per request — but at high volumes, these small per-request costs still add up.

The hardest part of estimating Azure OpenAI costs is knowing how many tokens each request uses. Here are typical ranges to help you calibrate the calculator:

Averages can be misleading when usage is skewed — a single customer running long-context RAG queries can use more tokens than a hundred customers running classification tasks. Estimate costs for your heaviest use case separately rather than averaging across all request types. The calculator above lets you model each use case individually, and you can sum the results for a more accurate total.

If you're already running Azure OpenAI in production, MarginDash can track your actual token usage per customer and show you exactly what each customer costs.