What Microsoft Foundry is
Microsoft Learn now titles the product Microsoft Foundry and records the lineage plainly: Azure AI Studio became Azure AI Foundry, which became Microsoft Foundry, while the old Azure AI Services became Foundry Tools (checked September 2026). It unifies agents, models and tools under one Azure resource with tracing, monitoring, evaluations, role-based access control, networking and policy. Three things get built on it: agents, either declarative prompt agents Microsoft hosts for you or hosted agents running your own containerised code; model deployments from a catalogue of more than 10,000 models including every Azure OpenAI model, Meta, Mistral, DeepSeek, Cohere, xAI and Microsoft's own; and tools, memory and retrieval through the Foundry Toolbox. Developers work in the portal at ai.azure.com, SDKs, the Azure Developer CLI or Visual Studio Code.
How it relates to Copilot Studio and Agent 365
Copilot Studio is, in Microsoft's words, a graphical low-code studio for building agents and workflows, aimed at makers and billed in Copilot Credits. Foundry is the developer platform underneath: code-first, model-agnostic and billed as Azure consumption. They meet at publishing — Microsoft Learn notes Foundry agents can be shared through Microsoft 365 Copilot, Teams and the Entra Agent Registry, so a Foundry-built agent can sit beside a Copilot Studio one in the same tenant. Agent 365 is neither: it is the governance layer that registers agents, gives them Entra identities and audits what they do, whichever tool built them. The practical split Frontrow uses: business-owned agents on Copilot Studio, engineering-owned agents on Foundry, and Agent 365 over both once they leave pilot.
The pricing model
There is no per-user licence. Models sold directly by Azure are billed through your Azure subscription, and Microsoft's deployment-type guidance describes pay-as-you-go processing on Global Standard and Data Zone deployments, with provisioned throughput as the alternative for workloads that need reserved capacity and predictable latency. Global Standard is where new models launch first, at the lowest price, with the broadest region coverage; Microsoft's advice is to start there and move only for a specific reason such as data residency or batch processing. Azure publishes Australian-dollar rates per model on its Microsoft Foundry pricing page, so budget with the pricing calculator against expected tokens rather than seats — and set Azure cost alerts before the first agent goes live.
Australian data residency — read the deployment type
Microsoft's rule is the same for every deployment type: data at rest stays in the designated Azure geography, but where inference happens depends on the type. Global deployments may process in any Azure region, Data Zone deployments stay inside a Microsoft-defined zone (Asia Pacific for Australian tenants) and Standard single-region deployments process in the deployment region only. Microsoft's region availability page (checked September 2026) lists Australia East for the GPT-5 family through GPT-5.6, GPT-4.1, GPT-4o, o-series and embeddings on Global Standard, but the single-region Standard list for Australia East is short — GPT-4.1 mini, GPT-4o and the text-embedding models — with a longer Regional Provisioned list including GPT-5, 5.1, 5.2, 5.4 and o3. If a contract demands in-country processing, that shorter list is your menu.
Frontrow's read for Australian buyers
Reach for Foundry when you need a model Copilot Studio cannot give you, a custom retrieval pipeline, an agent that runs your own code, or evaluations and tracing an auditor will accept. Reach for Copilot Studio when the agent is a workflow over Microsoft 365 data that a business team will own. Decide the residency question before the model question: if the answer is 'must process in Australia', design around the Australia East Standard or Provisioned lists above and accept that the newest models arrive there last. Then treat the Azure bill like any other consumption service — quotas per project, cost alerts and an owner — because the cheapest Foundry project is the one someone is watching.