Microsoft Foundry

What it is

Microsoft Foundry is Azure’s platform for calling models and building agents inside your own Azure subscription. It was Azure AI Studio, then Azure AI Foundry; the “Microsoft Foundry” name took over at the turn of 2026. Two layers matter to an ops buyer. Foundry Models is the catalog: OpenAI’s GPT-5.6 family (Sol, Terra and Luna, generally available since 9 July 2026), Claude (generally available in June 2026, with Anthropic operating inference as data processor), Microsoft’s own MAI models, Grok, Llama, DeepSeek and a long tail of open-weight models. Foundry Agent Service is the agent layer, and it now covers three shapes:

• Prompt agents — declarative agents (model, instructions, tools) that Microsoft runs for you. Generally available since 16 March 2026.
• Hosted agents — your own code (Microsoft Agent Framework, LangGraph or another open-source framework) running in Foundry-managed, isolated containers with session state and a filesystem. Generally available since 9 July 2026.
• Workflows and Routines — multi-step and scheduled agents; Routines went generally available on 24 September 2026.

Around those sit Toolboxes (one governed bundle of tools, MCP servers and connectors that many agents share; GA, with Skills and Tool Search still in preview), Foundry IQ (managed knowledge bases over SharePoint, OneLake, Azure AI Search and the web; knowledge bases GA, the serverless tier in preview), Memory (session, user and procedural memory, still preview), a browser automation tool (GA 19 October 2026) and bring-your-own AI gateway (GA, April 2026), which lets agents call models that sit behind Azure API Management or a third-party gateway. The newest piece, the Copilot harness, is a managed agent loop with skills and subagents that entered public preview on 28 October 2026.

Why ops teams end up here

The trigger is predictable: a team built its first agents in Copilot Studio, they worked, and now the credit bill or the ceiling hurts. A contract-intake agent needs custom extraction code. A pipeline-hygiene agent runs 3,000 times a month and the Copilot Credits for each action add up. The recruiting team wants Claude for screening rubrics but the rest of the stack is Microsoft. Foundry is where those agents move when an engineer takes ownership.

The reason to stay on Microsoft rather than jump to AWS is distribution and governance. Since 10 June 2026 a Foundry agent can be published straight into Microsoft 365 Copilot and Microsoft Teams, so the code-first agent shows up where employees already work, under the same Entra identity. Every agent gets an Entra Agent ID and registers in Microsoft Agent 365, which means security reviews the Foundry agent with the same Conditional Access, Purview and Defender tooling it uses for Copilot Studio agents. Spend draws down an existing Microsoft Azure Consumption Commitment (MACC), including Claude usage, which is often what clears procurement.

Pricing reality

There is no seat price. Prompt and workflow agents carry no runtime charge; you pay for model tokens and for the tools they call. Everything below is Azure list price in US East 2, checked against the Azure Retail Prices API on 4 November 2026.

• GPT-5.6 Terra: $2 per million input tokens and $12 per million output tokens on Global Standard. GPT-5.6 Sol is $4/$20; GPT-5.6 Luna is $0.20/$1.20. These are short-context rates; long-context prompts cost more.
• Data Zone deployments cost 10% more. Keeping inference inside the EU or US data zone turns Terra into $2.20/$13.20. Batch is 50% off Global Standard.
• Claude bills in Claude Consumption Units on your Azure invoice at Anthropic’s published token rates, and counts toward MACC. Anthropic, not Microsoft, is the data processor.
• Hosted agents: $0.0994 per vCPU-hour and $0.0118 per GiB-hour of container compute.
• Memory: $0.25 per 1,000 short-term events stored, $0.25 per 1,000 long-term memories a month, $0.50 per 1,000 retrievals.
• Tools bill separately: Grounding with Bing, Foundry IQ (as Azure AI Search), file search storage, code interpreter sessions, Logic Apps connectors and browser automation (as Playwright Workspaces) are each their own meter.
• Copilot harness: free in preview until 1 December 2026, then the same $0.0994 per vCPU-hour and $0.0118 per GiB-hour for its sandbox.

A worked band: an internal RevOps agent handling 3,000 sessions a month, each about 20,000 input and 1,500 output tokens on GPT-5.6 Terra, pinned to the EU data zone. Tokens: about $132 input plus $59 output, roughly $191. If it is a prompt agent, that is nearly the whole bill. If it is a hosted agent holding one vCPU and 2 GiB for five minutes per session, add about $31 of compute. Memory and grounding together stay under $15 at this volume. Total: roughly $200 to $240 a month, about $2,400 to $2,900 a year. The same workload as a Copilot Studio autonomous agent at 6 actions and one generative answer per run is about 96,000 credits, or $800 a month in capacity packs. The Azure bill is the small number; the engineer who builds and maintains the agent is the real cost.

Best for

A platform or RevOps engineering team in a Microsoft 365 and Azure shop that has outgrown Copilot Studio’s low-code ceiling or its per-action credits, and needs agents that ship into Teams and Microsoft 365 Copilot under Entra identity and Agent 365 governance. It wins outright when the workload is high-volume and back-office (extraction, classification, routing, enrichment) and the company already has an Azure commit to burn down.

Not for

A 10-50 person ops team with no engineer to own the build: stay in Copilot Studio or buy Dust or Langdock. Foundry gives you a portal and SDKs, not an assistant your RevOps manager can configure on their own. It is also the wrong buy for a company whose identity and data live in AWS or Google Workspace; you would be adopting Entra, Azure subscriptions and a second cloud bill to get what your existing cloud already sells.

Versus the alternatives

• Amazon Bedrock — the other market-share leader among hyperscaler agent platforms. Pick Bedrock and AgentCore when identity, data and the cloud commit are in AWS. Pick Foundry when they are in Entra and Azure, or when the agent has to appear inside Teams and Microsoft 365 Copilot. Token prices are close; the deciding factor is where your identity provider and data already live. The model gap has also closed: GPT-5.x runs on Bedrock and Claude runs on Foundry.
• Google Cloud’s Agent Platform — the third hyperscaler option. Pick it when the warehouse is BigQuery, the company runs on Google Workspace and Gemini is the first-choice model.
• Copilot Studio — Microsoft’s own low-code layer above Foundry. Keep agents there while makers, not engineers, own them and volume is low enough that credits stay under a few packs a month. Move to Foundry when you need custom code, a model Copilot Studio does not expose, or per-token pricing at volume.
• Claude Platform and Claude Managed Agents — the fastest-moving entrant for agent features. Claude in Foundry lacks code execution, Agent Skills, the Files API and batches, which Anthropic’s own platform has. Pick Anthropic directly when those features drive the design; pick Foundry when the requirement is one Azure invoice and Entra-governed distribution.

Watch-outs

• Claude in Foundry is not the full Claude API, and Microsoft is not its processor. Anthropic operates inference and is the data processor and SLA provider, and the hosted version lacks code execution, Agent Skills, the Files API and batches. Guard: have privacy counsel review Anthropic’s terms as a separate processor before routing personal data to Claude, and check every Claude feature your design needs against Microsoft’s Claude capability list before you commit.
• Data residency is a deployment type, not a default. Global Standard is the cheapest and can route a request to any region. Guard: for EU or regulated data, deploy only Data Zone or Regional SKUs, budget the 10% premium, and enforce it with an Azure Policy that denies Global Standard deployments in the subscriptions that hold those agents.
• GA and preview are mixed inside one agent. Hosted agents and Toolboxes are GA, while Memory, Toolbox Skills, Tool Search, serverless Foundry IQ and the Copilot harness are still preview, without GA commitments. Guard: keep a dependency list per agent that marks each feature GA or preview, and keep preview features out of anything customer-facing or compliance-relevant until they ship GA.
• The bill arrives on six meters. Tokens, hosted compute, memory, Bing grounding, AI Search and Playwright each bill separately, so no single view shows what one agent costs. Guard: put each production agent in its own resource group with an agent-name tag, and set an Azure Cost Management budget alert per agent at 120% of forecast.
• Older tutorials teach retired APIs. The Assistants API sunset on 26 August 2026 and Prompt flow retires on 20 April 2027. Guard: build only on the Responses API and Foundry Agent Service, and move any Prompt flow pipeline onto Agent Service workflows before you add features to it.

Related: the enterprise AI rollout stack and the AI agent ops stack show where a model and agent platform like Foundry sits next to the assistant, governance and observability layers.