What Is Azure AI Models? A Plain-English Guide for 2026

Azure AI Models Has a New Name (And Why That Confuses Everyone)

You type "Azure AI models" into a search bar in 2026, and the results send you in circles. One tab says "Azure AI Foundry Models," another says "Foundry Models," and the official documentation now calls it Microsoft Foundry Models. The confusion is real, and it costs you time before you even start evaluating anything.

Here's the mapping: the offering you remember as Azure AI Models is now branded Microsoft Foundry Models, described by Microsoft Learn as "your one-stop destination for discovering, evaluating, and deploying powerful AI models." It is not a single model. It is a catalog, and the catalog is large.

That catalog holds 11,000+ models from providers including OpenAI, Anthropic, Meta, Mistral, DeepSeek, Cohere, NVIDIA, xAI, and Hugging Face, according to Microsoft's product page. Microsoft Learn puts the count at "over 10,000 models, with approximately 50 new models published each month." The two figures point in the same direction; the catalog is expanding faster than most documentation can keep up with.

The scale behind it matters. The same product page reports 80,000+ enterprises using Foundry, including 80% of the Fortune 500. Microsoft's Q2 FY2026 earnings, covered by Futurum Group, noted over 1,500 customers running both Anthropic and OpenAI models through the platform, and more than 250 customers on track to process over 1 trillion tokens on Foundry this year. This is not a side experiment. It is core enterprise infrastructure.

The catalog splits into two tiers that matter for selection:

• Foundry Models sold by Azure: Microsoft-hosted, with enterprise SLAs and deep Azure integration.
• Foundry Models from partners and community: the broader catalog, supported by external providers.

Microsoft's own guidance: Azure-sold models suit scenarios needing deep integration, guaranteed support, and enterprise SLAs; partner and community models suit specialized, innovation-led use cases.

For readers building agents on top of these models, the catalog sprawl creates a downstream problem. Every model choice is also a stack choice, and agent skills can extend what those models actually do in production. SkillzBench is building a free directory and MCP connector where skills are version-pinned, security-scanned, and benchmarked for Lift, so you can test which ones actually help before wiring them into an Azure-hosted agent.

Inside the Catalog: Two Tiers, Six Model Types, and How to Choose

Eleven thousand models sounds like a blessing until you have to pick one. The good news: Microsoft sorts the Foundry Models catalog into two tiers and six model types, and understanding that structure cuts most of the noise.

The two tiers

According to Microsoft's own documentation, the catalog splits into "Foundry Models sold by Azure" — Microsoft-hosted, with enterprise SLAs, guaranteed support, and deep Azure integration — and "Foundry Models from partners and community," which make up the vast majority of the catalog and are supported by external providers. Microsoft's guidance is refreshingly direct: Azure-sold models suit scenarios requiring deep integration and enterprise SLAs; partner and community models "excel in specialized use cases and innovation-led scenarios."

The scale here is real. The product page lists 11,000+ models from providers including OpenAI, Anthropic, Meta, Mistral, DeepSeek, and xAI, with roughly 50 new models published each month. Over 1,500 customers already run both Anthropic and OpenAI models through Foundry, per Microsoft's Q2 FY2026 earnings coverage.

The six model types

• Foundation models — general-purpose workhorses for broad language tasks
• Reasoning models — built for complex, multi-step problem-solving
• Small language models — compact, cost-efficient options like Phi, which has passed 60 million downloads on Foundry
• Multimodal models — handle text, images, and other inputs together
• Domain-specific and industry models — tuned for verticals like finance, healthcare, and legal

Practical selection rules

For frontier work, the 2026 lineup offers GPT-5.2 and Claude 4.5, both available through the Azure-sold tier. For cost-sensitive, high-volume tasks, small language models like Phi are the sensible pick — Phi-3-mini reportedly beats models twice its size on benchmarks while supporting 128K-token contexts. And if data residency is your constraint, regional and sovereign models from Mistral and Cohere, deployed into a chosen region rather than a shared endpoint, address that directly.

One caveat worth remembering: a third-party practitioner comparison argues that frontier models "land within a few points of each other on most production tasks" — so don't agonize over model choice at the expense of the deployment machinery underneath it. If you're building agents on these models, agent skills can automate much of the surrounding workflow — SkillzBench is building a free, security-scanned directory of exactly those skills, with benchmarks to show which ones actually lift performance.

Paying for AI Models: Serverless vs. Managed Compute (and the Hidden 15-40%)

The sticker price of an AI model is rarely the price you pay. Once you move from a demo to production, the deployment path you choose — and the costs vendors don't put on the pricing page — determine whether your AI budget survives contact with reality.

Azure gives you two paths. Serverless pay-as-you-go bills per token: GPT-5 runs at $1.25 per million input tokens and $10 per million output tokens, with a 50% discount if you can tolerate batch processing, per practitioner analysis of enterprise AI platforms. Managed compute gives you dedicated VMs billed per core hour — predictable, but you pay whether or not anyone queries the model.

The trap is that token rates aren't your real bill. The same analysis found production deployments commonly add 15% to 40% over published token rates, driven by costs that hide in the fine print:

• Data egress — moving responses out of the region where the model runs
• Private Link — the networking layer most enterprises require for compliance
• Log Analytics ingestion — the telemetry that makes AI auditable
• Support plans that add a percentage on top of usage

Two more realities deserve attention. First, the quota-versus-capacity problem: holding PTU quota means you're permitted to deploy that many units, but if the region lacks hardware at deployment time, the deployment fails. Second, the standard pay-as-you-go endpoint carries no latency promise — Azure OpenAI offers a financially backed 99.9% monthly availability SLA, but during peak demand you can see steady 429 throttling that never breaches the uptime metric. You paid full price for degraded service with no credit recourse.

This is why experienced teams don't commit to reserved capacity from a calculator. The recommended discipline is to run pay-as-you-go for 30 to 60 days, measure P95 hourly token throughput, then model the reserved option against real numbers. Microsoft's own adoption data suggests scale matters here: over 250 customers are on track to process more than 1 trillion tokens on Foundry this year, and customers spending more than $1M per quarter grew nearly 80% year over year, per earnings-based coverage. At that volume, the 15–40% overhead compounds fast.

One way to trim the bill regardless of deployment path is to make the agent layer itself more efficient. Well-chosen agent skills can cut wasted tokens by handling routing, formatting, and validation before requests hit the model — and SkillzBench is building a free directory where skills are version-pinned, security-scanned, and benchmarked, so you can test which ones actually reduce work rather than add it.

Budget for the machinery, not just the model. The token rate is where the conversation starts, not where it ends.

From Model to Agent: Building on Foundry Without Getting Burned

Picking a model from a catalog of 11,000+ options feels like the big decision. It isn't. Practitioner analysis of the major cloud AI platforms puts it bluntly: frontier models "land within a few points of each other on most production tasks," and the decision that sticks is the machinery underneath the model.

On Azure, that machinery increasingly means agents. Agent work routes through the Foundry Agent Service — the supported replacement for the retired Assistants API — so the question shifts from "which model?" to "how does my agent integrate, where does its data live, and who governs it?" Those choices are hard to undo later.

They're also where Azure differentiates itself. Models deploy into a region you choose rather than a shared endpoint, which is what gives Azure its data-residency story. Microsoft is doubling down, with sovereignty-ready deployments across public, private, and national partner clouds and new data center commitments in seven countries, according to Q2 FY2026 earnings coverage. For regulated industries, that often matters more than a point or two on a benchmark.

The operational caveats deserve equal attention before you ship:

• Quota is not capacity: holding PTU quota doesn't guarantee hardware exists in your region at deployment time.
• Production overheads commonly add 15–40% over token rates through support plans, egress, Private Link, and Log Analytics.
• Standard pay-as-you-go endpoints carry no latency promise — throttling during peak demand never breaches the 99.9% availability SLA.
• Run pay-as-you-go for 30–60 days and measure P95 throughput before committing to reserved capacity.

There's one more layer agents add: skills. Agent skills — instruction packages that extend what an agent can do — are proliferating fast, and published research found roughly 26% of them carry at least one vulnerability. If you're extending Azure-hosted agents with third-party skills, you inherit a trust problem that has nothing to do with Microsoft's infrastructure.

That's the gap SkillzBench is building for. It's a free, independent directory of agent skills where every skill is version-pinned to a commit SHA, security-scanned, license-checked, and benchmarked in a sandbox to show its Lift in percentage points. A free MCP connector (in build, with the waitlist live) will let you toggle exactly which skills each agent can see — so the extension layer gets the same scrutiny as the platform underneath it.

Choose the model lightly. Choose the machinery, the region, and the skills deliberately — because those are the decisions you'll live with.