Buying OpenAI and Claude Tokens Through Microsoft Foundry

A company that wants to wire GPT or Claude into its own software has two routes. One is to open separate accounts with OpenAI and Anthropic and put a credit card on each. The other is to use both vendors’ models from inside an Azure subscription, through Microsoft Foundry (formerly Azure AI Foundry). On the second route you never buy a token pack. Whatever you use is added to your Azure invoice at the end of the month. The two model families reach that invoice by different paths, and the difference decides which subscription you should start with.

Tokens are metered, not purchased

A token is the unit a model uses to measure the text it reads and writes. Your prompt counts as input, the model’s answer counts as output, and the two have separate unit prices. Foundry holds no prepaid balance for them. You create a model deployment, your application sends requests to it, Azure meters the usage and bills it monthly.

Microsoft splits its model catalog into two groups. OpenAI models sit under “Foundry Models sold by Azure”. Claude models sit under “Foundry Models from partners and community”. The group a model belongs to determines how its token charges arrive on your bill.

OpenAI models bill on native Azure meters

Microsoft hosts and sells GPT models under its own Product Terms. Usage is measured by an Azure meter, the same way a virtual machine or a storage account is, and no Marketplace subscription is involved. How you pay depends on the deployment type:

• Standard. You pay per token. It suits workloads whose traffic you cannot predict.
• Provisioned (PTU). You reserve a fixed amount of processing capacity at an hourly rate. The meter runs even if you send no requests. A one-month or one-year reservation lowers the hourly rate.
• Batch. You submit requests in bulk as a file. The target turnaround is 24 hours and the price is half of Global Standard.

Microsoft recommends starting with Global Standard for most workloads, and its documentation gives the reasons. New models launch there first, and it has the lowest price and the broadest region coverage. PTU starts to make sense once traffic is predictable and consistently high. According to the Azure pricing page the minimum deployment for GPT-6 Sol and Astra is 15 PTUs, which is a large commitment for a pilot.

Current per-token prices are listed model by model on the Azure OpenAI pricing page and in the Azure pricing calculator.

Claude models arrive through Azure Marketplace

Anthropic is the seller of Claude models. Microsoft issues the invoice and handles support requests. The first time you deploy a Claude model you accept the “Claude Platform on Foundry” offer in Azure Marketplace, which requires the Owner or Contributor role on the subscription.

Usage then appears on the invoice in a single unit called the Claude Consumption Unit (CCU). A CCU is easy to mistake for prepaid credit. Microsoft’s documentation says it is a billing unit only and that there is no CCU balance to buy in advance. Anthropic rates the input and output tokens of each call in dollars at its list prices, applies any contractual discount, converts the result at 0.01 dollars per CCU and reports it to Azure Marketplace hourly. The invoice is issued monthly in arrears.

Anthropic’s price list on 11 November 2026 shows Claude Sonnet 5.5 at 2 dollars per million input tokens and 10 dollars per million output tokens. An application that consumes 50 million input and 10 million output tokens in a month pays 100 dollars for input and 100 dollars for output. The 200 dollar total shows up as 20,000 CCU. The amount is calculated in dollars and converted to the currency of your Azure invoice.

Claude has fewer deployment types. The documentation lists Global Standard and US Data Zone Standard only, with no PTU and no Batch. Choosing the US Data Zone applies a 1.1x multiplier to token prices.

Claude cannot be deployed on a CSP subscription

If you buy Azure from a partner under the Cloud Solution Provider (CSP) program, this section matters to you. Microsoft’s documentation lists the subscription types that Claude models do not support:

• CSP subscriptions
• Free trial and student subscriptions
• Sponsored subscriptions that run on credits only
• Enterprise accounts located in South Korea

OpenAI models fall outside this restriction because they bill on Azure’s own meters. The same document advises customers who hit the CSP error to use models sold directly by Azure instead. With a CSP subscription you can deploy GPT models today. Claude needs a separate subscription bought directly from Microsoft with a pay-as-you-go payment method on file.

Country is not an obstacle for customers in Türkiye. Türkiye is on the list of countries where the Anthropic offer is available for purchase.

Where is the data processed?

Buying tokens from Azure does not mean your data stays in your own country. A Global deployment can process a request in any Azure region where the model runs. A Data Zone deployment limits processing to the US, the EU or Asia Pacific. Models sold by Azure offer an EU Data Zone option, and the region availability table shows which models support it. For Claude, Data Zone exists for the US only, and Sweden Central is the single European region where Claude can be deployed.

Claude also comes in two hosting versions. With “Hosted on Azure”, prompts and outputs never leave Azure infrastructure. With “Hosted on Anthropic infrastructure”, data may be processed outside Azure. Some of the newest models are offered only in the second version and in preview, Claude Fable 5.2 among them. Whichever version you pick, Anthropic is the data processor and Anthropic’s commercial terms and data processing addendum apply.

If requests will contain personal data, have your legal counsel confirm which deployment type satisfies the cross-border transfer rules of KVKK, Türkiye’s personal data protection law, or of the regime you operate under. Write down who may send which data to which model in your corporate AI policy.

How do you cap the spend?

On pay-as-you-go, a faulty loop or unexpected traffic can inflate the bill overnight. Put two safeguards in place on day one. Define a monthly budget with threshold alerts on the Foundry resource in Azure Cost Management. Keep each deployment’s tokens-per-minute quota as low as your workload allows, because the quota also caps what can be spent per minute.

Per-model usage is visible on the Monitor tab of the Foundry portal. Claude has one limitation worth knowing. Resource tags do not give project-level cost attribution, and charges are recorded against the Foundry resource. If you need to see several projects separately, give each project its own Foundry resource. Our Azure pricing guide covers how to read the rest of the invoice.

Which route should you start with?

The decision comes down to three questions.

If your subscription is CSP and GPT models do the job, start with a Global Standard deployment in the subscription you already have. No extra contract, invoice or account is needed.

If Claude is a requirement, you have two options. Open a separate subscription directly with Microsoft and use Claude through Foundry, or get an API account from Anthropic. On the Foundry route billing and support stay with Microsoft, and the spend counts toward your Microsoft Azure Consumption Commitment (MACC) if you have one. In my view, a company that already runs infrastructure on Azure is better off with Foundry than with a second vendor account to manage.

Leave PTU off the table for the first months. Measure real consumption on a standard deployment for a month or two, then size the reservation with that data.

If you would like us to map which models you can deploy with which deployment type on your subscription, get in touch.

Sources

• Understanding deployment types in Microsoft Foundry Models, Microsoft Learn
• Provisioned throughput for Foundry Models, Microsoft Learn
• Claude consumption units (CCU) billing in Microsoft Foundry, Microsoft Learn
• Compare hosting options for Claude models in Microsoft Foundry, Microsoft Learn
• Foundry Models from partners and community, Microsoft Learn
• Pricing, Claude Platform Docs
• Azure OpenAI Service pricing