Winmill helps teams ship reliable, compliant AI. In this post, we explain Microsoft Foundry Models. We’ll discuss how they differ from (and include) Azure OpenAI, and a practical transition plan for organizations standardizing on the new Foundry platform.
Overview
- Microsoft Foundry Models is the unified, enterprise catalog and runtime for discovering, evaluating, and deploying models from Microsoft, OpenAI, Anthropic, Mistral, Meta, DeepSeek, xAI, Black Forest Labs, and more—now exceeding 11,000+ options with multiple deployment patterns and governance built in.
- Foundry provides “models sold directly by Azure” (including Azure OpenAI) plus partner/community models, with default safety policies and configurable guardrails.
- If you use Azure OpenAI today, you can upgrade resources to the Foundry resource type while keeping endpoints, security, and fine‑tuning state—gaining agents, model router, broader catalog, and governance.
What are “Microsoft Foundry Models”?
Foundry Models is Microsoft’s model marketplace and inference layer inside Microsoft Foundry (formerly Azure AI Foundry). It centralizes model discovery → evaluation → deployment across OpenAI and non‑OpenAI providers, with serverless APIs, managed compute, Provisioned Throughput reservations, and integrated governance.
- Catalog breadth: Foundry lists models from Microsoft/OpenAI alongside Anthropic (Claude), Mistral, Meta Llama, DeepSeek, xAI (Grok), Black Forest Labs (FLUX), Cohere, and more.
- Operational features: built‑in Model Router, evaluation/leaderboards, priority processing, and fine‑tuning—designed for multi‑model strategies and cost/performance optimization.
- Safety and trust: default Content Safety and guardrails apply to prompts/outputs (hate/violence/sexual/self‑harm, protected material, jailbreak detection). You can tune thresholds or attach custom policies per deployment.
Bottom line: Foundry Models is the model layer of Microsoft’s unified AI platform. You can still run your Azure OpenAI workloads—but now you can evaluate and switch across a much wider set of models from one place.
Azure OpenAI vs. Microsoft Foundry: what’s changing?
Many teams ask whether Foundry replaces Azure OpenAI. The answer: Foundry subsumes and extends Azure OpenAI. You retain direct access to OpenAI models, but also get the broader multi‑model ecosystem, agents, and governance in one platform.
Why Winmill recommends standardizing on Foundry Models now
- Choice without rewrites. The OpenAI SDK / v1 endpoints work against Foundry endpoints, and Microsoft provides migration guidance for inference SDK users—making it straightforward to adopt the catalog and keep your code patterns.
- Enterprise guardrails baked in. Default guardrails + configurable safety allow you to enforce policy consistently across providers and projects—crucial for policy and audit teams.
- Futureproofed roadmap. Microsoft is investing heavily: 11,000+ models, Model Router GA, agent service, and expanded partnerships announced at Ignite 2025.
Architecture at a glance
Core components you’ll use on day one:
- Foundry Models (catalog + endpoints) for model selection and deployment.
- Default Guardrails (content filters, protected‑material checks, jailbreak detection), configurable per model.
- Model Router to auto‑route prompts to best‑fit models and roll back if a new model underperforms.
- Foundry Agent Service when you need tool use, multi‑agent patterns, and enterprise runtime (identity, networking, observability).
Azure OpenAI vs Microsoft Foundry transition
Here’s a pragmatic method to migrate without breakage.
A. Upgrade paths
- Resource upgrade (recommended): Convert an existing Azure OpenAI resource to a Foundry resource. This preserves your endpoint, keys, network config, and fine‑tuning state, while unlocking the broader Foundry catalog and agents.
- Project migration: For legacy hub‑based projects, move to Foundry projects to access the latest APIs and consolidated governance.
- Agents migration: Use the migration tooling to move from the older Assistants API to Foundry Agents for better orchestration, storage options (incl. BYO Cosmos DB), and multi‑agent workflows.
B. API & SDK considerations
- Prefer the OpenAI SDK (v1 endpoints) for broadest cross‑model compatibility across Azure OpenAI and Foundry Models. Microsoft’s guide explains auth, base URLs, and parameter differences vs. the Azure AI Inference SDK.
C. Safety & governance parity
- Re‑apply your Content Safety thresholds and blocklists at the Foundry deployment layer; defaults cover harms, protected material, and jailbreak detection.
Implementation Checklist
- Inventory your models & endpoints. Capture which GPT deployments you have today and their quotas; note any fine‑tunes and batch jobs. (You’ll keep these during the upgrade.)
- Upgrade one Azure OpenAI resource to Foundry in a lower environment; validate endpoint continuity and RBAC.
- Attach guardrails: keep defaults at “medium” severity for harms, enable protected‑material detectors, and test jailbreak detection with your prompts.
- Pilot Model Router on one workload to compare cost/quality vs. fixed‑model routing; capture metrics in observability.
- (Optional) Introduce Agents for tool‑calling scenarios and multi‑step workflows once endpoints and guardrails are stable.
FAQ
Is Microsoft Foundry just a rebrand?
It’s more than a rename: Foundry consolidates models, agents, tools, and governance in a platform designed for multi‑model and agentic workflows—not just OpenAI access.
Can I still call GPT models directly?
Yes. Azure OpenAI remains first‑class inside Foundry. Upgrading resources lets you keep your existing endpoint while adding the broader catalog and features.
How big is the model catalog now?
Microsoft reports over 10,000–11,000+ models (including open‑source) with expanding partner coverage; the catalog front‑end surfaces both Microsoft‑sold and partner/community models.
What about safety defaults?
Default guardrails apply across models (text, vision) with configurable thresholds and detectors; you can attach custom filters per deployment.
Where Winmill plugs in
- Data & RAG engineering for better answer quality, citations, and groundedness—see our Data Intelligence & AI services.
- Rapid prototyping (MVP) in Foundry—we deliver a working AI app or agent with guardrails and observability in weeks, not months.
Modernize Your AI Stack With Zero Disruption
Run a guided migration sprint with Winmill. We’ll upgrade your Azure OpenAI workloads to Foundry, enable default guardrails, pilot Model Router, and stand up observability—so you can scale multi‑model AI with confidence.
- Keep your existing endpoints, keys, and fine‑tunes
- Default + custom guardrails configured to your policy
- Model Router pilot with cost/quality benchmarks