What is Microsoft Foundry?
Microsoft Foundry, formerly Azure AI Studio and then Azure AI Foundry, is Azure's platform for building with AI models and agents. A single Foundry resource holds projects, model deployments, agents and tools under one set of Azure access controls, networking and policies. Microsoft's documentation lists more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others.
Foundry Agent Service runs agents in three forms: prompt agents you configure in the portal, voice agents through Voice Live, and hosted agents that run your own container code built with Microsoft Agent Framework, LangGraph or Semantic Kernel. Toolboxes give agents one managed endpoint for tools, MCP servers and enterprise data. Browser Automation reached general availability on 19 October 2026, and on 28 October Microsoft added the Copilot harness, the agent loop behind GitHub Copilot and Copilot Studio, in public preview.
The platform itself is free; you pay for model tokens, hosted agent compute and the tools you attach.
How much does Microsoft Foundry cost?
| Plan | Price | What you get |
|---|---|---|
| Hosted agents | $0.0994 per vCPU-hour | Plus $0.0118 per GiB-hour of memory |
| Grounding with Bing | $14 per 1,000 transactions |
Platform free; models billed per token at each model's rate. Hosted agents $0.0994 per vCPU-hour and $0.0118 per GiB-hour of memory (East US); Grounding with Bing $14 per 1,000 transactions; file search 1 GB free. Prompt agents carry no extra charge beyond model and tool use.
Is Microsoft Foundry worth it?
Most companies do not want a separate contract, login and data agreement for every AI model their engineers try. They want one bill, their existing security rules and a promise that prompts stay inside their cloud.
Foundry is Microsoft's answer. One Azure resource gives you OpenAI's GPT models, Anthropic's Claude, Microsoft's own MAI models and thousands of open models, all under your Entra ID sign-in, private networking and Azure Policy. The Model Router now picks among 35 models per request, and Anthropic models must be deployed to your resource before it can use them.
The data terms are the strongest selling point. Microsoft's documentation says prompts and outputs for models sold by Azure are not shared with OpenAI or other model makers and are not used to train foundation models. Data stored for features like the Responses API stays in your chosen geography, encrypted, and you can delete it.
The agent side has matured quickly. Agent Service reached general availability in March 2026, hosted agents and Toolboxes followed in the summer, and scheduled Routines and agent-to-agent protocols went live in September. Hosted agents cost $0.0994 per vCPU-hour plus $0.0118 per GiB-hour of memory in East US, and prompt agents add no charge beyond the model and tools they use.
The catch: Foundry is a lot of Azure. You need a subscription, a resource, role assignments and often a quota request before your first agent runs. Microsoft is midway through replacing the classic portal, hub-based projects and the Assistants API with new equivalents, so tutorials older than a year often point at the wrong screen. Several headline features, including the Copilot harness and agent memory, are still in preview, which Microsoft says carries no service-level agreement.
Watch the data defaults too. Global and Data Zone deployments, usually the cheapest and fastest, may process your prompts in any region where the model runs. Abuse monitoring can store flagged prompts for review by Microsoft staff unless you apply for, and are granted, modified monitoring.
Our take: Foundry is less a product than the AI wing of an Azure estate; teams already inside that estate get the best data terms in the market, and everyone else gets a long checklist before the first token.
Where it earns its place
A platform team at a 2,000-person bank on Azure
Give developers GPT and Claude without new vendor contracts
Deploy both into one Foundry resource behind private networking and Entra ID; usage lands on the existing Azure invoice, and Data Zone deployments keep processing inside the EU for EU resources.
An engineer shipping an internal agent built with LangGraph
Run agent code in production without managing containers and scaling
Package it as a hosted agent; Foundry provides the endpoint, scaling, identity and tracing. One vCPU with 2 GiB running all month costs about $90 in East US before model tokens.
A support operations lead at a telecoms company
Let an agent check order status on supplier websites that have no API
Attach Browser Automation, generally available since 19 October 2026, through a Toolbox; it runs in a separate Playwright Workspace with a live view so a person can take control.
Before you commit
- Setup
- Create an Azure subscription and a Foundry resource, assign roles, deploy a model and call it from the portal playground or the azure-ai-projects SDK; Azure OpenAI resources can be upgraded in place, keeping endpoints and keys.
- Integrations
- SDKs for Python, C#, JavaScript and Java, the Azure Developer CLI, a VS Code extension, a Foundry MCP server and skill for coding agents, and publishing to Microsoft Teams and Microsoft 365 Copilot.
- Data and privacy
- Prompts and outputs for models sold by Azure are not shared with model providers or used to train foundation models; stored data stays in your geography, encrypted with AES-256 or your own keys. Flagged content may be reviewed by Microsoft staff unless modified abuse monitoring is approved.
- Cost at scale
- The platform is free; costs are model tokens at each model's rate, hosted agent compute, Grounding with Bing at $14 per 1,000 transactions and Azure AI Search for retrieval. One-year Agent Commit Unit plans give 5% to 15% off.
- Limits
- Model quotas vary by region and deployment type and often need a request to raise; preview features carry no SLA, and Copilot harness agents must choose it at creation.
What’s good
- OpenAI, Anthropic, MAI and open models under one Azure bill and identity
- Prompts not shared with model makers or used to train foundation models
- Managed runtime for hosted agents, from $0.0994 per vCPU-hour
- Private networking, customer-managed keys and Azure Policy
What’s not
- Needs an Azure subscription, roles and often quota requests to start
- Copilot harness and agent memory still in preview with no SLA
- Global deployments can process prompts outside your region
- Old tutorials point at the retired classic portal and Assistants API
Best for
Engineering teams in organisations already standardised on Azure who need several model vendors under one contract and security model.
Microsoft Foundry alternatives
Microsoft Foundry: common questions
- Is Microsoft Foundry worth it?
- If your company already runs on Azure, Foundry is the sensible home for its AI work; if it does not, the setup tax is hard to justify. We rate it 4.3 out of 5.
- How much does Microsoft Foundry cost?
- Platform free; models billed per token at each model's rate. Hosted agents $0.0994 per vCPU-hour and $0.0118 per GiB-hour of memory (East US); Grounding with Bing $14 per 1,000 transactions; file search 1 GB free. Prompt agents carry no extra charge beyond model and tool use.
- What are the downsides of Microsoft Foundry?
- Needs an Azure subscription, roles and often quota requests to start; copilot harness and agent memory still in preview with no SLA; global deployments can process prompts outside your region; old tutorials point at the retired classic portal and Assistants API.
- Who is Microsoft Foundry best for?
- Engineering teams in organisations already standardised on Azure who need several model vendors under one contract and security model.
- How much does Microsoft Foundry cost for a team or in production?
- The platform is free; costs are model tokens at each model's rate, hosted agent compute, Grounding with Bing at $14 per 1,000 transactions and Azure AI Search for retrieval. One-year Agent Commit Unit plans give 5% to 15% off.
- Is Microsoft Foundry safe for company data?
- Prompts and outputs for models sold by Azure are not shared with model providers or used to train foundation models; stored data stays in your geography, encrypted with AES-256 or your own keys. Flagged content may be reviewed by Microsoft staff unless modified abuse monitoring is approved.
- What are the best Microsoft Foundry alternatives?
- The closest alternatives we have reviewed are Google Agent Platform (Vertex AI), Amazon Bedrock, OpenRouter and Google AI Studio.
Answers checked Nov 6, 2026.
Make Microsoft Foundry? Show this rating on your site
