What changed in the Azure AI Gateway labs for each product release, and where to practice every new feature. Newest first. 🆕 marks a new lab, ✏️ an updated lab, ✅ a lab that already covered the feature, and ➖ a feature that no lab covers yet.
November 18, 2026
Azure AI Gateway is generally available
Azure AI Gateway is generally available. Every lab now uses the generally available portal and management API
version 2026-11-01, and every track was retested on it.
Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Create a gateway, import a model, and make your first call in the generally available portal | Quick start · Your first model call in 5 minutes and 01 · Provision your AI Gateway | ✏️ Updated |
Manage gateways and their assets with the management API 2026-11-01 in REST and Bicep | Every lab's REST and Bicep + azd tracks, starting with 02 · Review the templates | ✏️ Updated |
Manage gateways and their assets with the az aigateway extension 1.0.0b34, which calls 2026-11-01 and prints the generally available field names | Every lab's Azure CLI track, starting with Install the aigateway extension, and 08 · Set the policy, which writes content safety with az aigateway model update --policies | ✏️ Updated |
Manage gateways and their assets with Terraform and the AzAPI provider at 2026-11-01 | A new Terraform track in the quick start and 28 labs, starting with Deploy a lab with Terraform and 01 · Provision your AI Gateway | ✏️ Updated |
Providers set modelProviderType and a displayName, and a model's supportedEndpoints decide its API format | 02 · Publish a Foundry model and make your first call and 03 · Publish models from any provider behind stable aliases | ✏️ Updated |
Tool server backends set mode (failClosed or failOpen), and tool lists use blockList | 04 · Federate MCP tools behind one endpoint and 05 · Control which tools agents can call | ✏️ Updated |
Request rate limits set a count and a period | 05 · Control which tools agents can call and 09 · Review the policy file | ✏️ Updated |
Cost limit overrides for Microsoft Entra identities list claim name and value pairs | 07 · Give a Microsoft Entra identity its own budget | ✏️ Updated |
| Token limits and content safety policies on the generally available API | 06 · Throttle tokens per caller with token limits and 08 · Screen prompts and responses with content safety | ✏️ Updated |
| Deploy the enterprise reference architecture with the generally available templates | 21 · Review the templates and 21 · Apply the gateway baseline | ✏️ Updated |
| Current agent SDKs and the coding-agent plugin | 15 · Connect agent frameworks to models and tools, 16 · Deploy a Foundry hosted agent behind AI Gateway, and 17 · Build on your gateway with coding agents | ✏️ Updated |
| Import Foundry deployments filtered by resource group | 24 · Import the deployment and 25 · Import the Claude deployments | ✏️ Updated |
Back a model provider with a system- or user-assigned identity (authentication.managedIdentity), or an API-key header | 12 · Point the model provider at the identity and 12 · Use stored credentials when a backend can't take a managed identity | ✏️ Updated |
Telemetry exporters set telemetryExporterType, {url} endpoints, and a top-level authentication block | 13 · Create the telemetry exporter and 14 · Choose the claims to export | ✏️ Updated |
| Configure telemetry in the portal on a managed Log Analytics workspace | 13 · Start the setup | ✏️ Updated |
Turn off a gateway's public network access: its public endpoint returns 403 | 18 · Turn off the gateway's public network access | ✏️ Updated |
| Request inbound private endpoints, find the gateway's private DNS zone, and reach the gateway on a private IP | 18 · Inbound private endpoints and 18 · Reach the gateway through an inbound private endpoint | ✏️ Updated |
| Protect tool calls with tool threat protection on the generally available tool server shape | 19 · Add tool threat protection | ✏️ Updated |
| Allow exact tool names by pattern, block tools, and suspend MCP server access | 20 · Allow the exact tool names and 20 · Block the whole server | ✏️ Updated |
November 12, 2026
Model failover and retry rules
Keep an alias available when a backend deployment is throttled or fails: retry the same backend, switch the request to the next eligible model, and temporarily exclude the failing backend from later requests. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Ordered failure rules on an alias, with a CEL condition over the backend response | 29 · Turn on failover | 🆕 New lab |
Switch a failed request to the next backend model by priority, and prove it with x-ms-served-model | 29 · Watch the failover | 🆕 New lab |
Temporarily exclude a failing backend (excludeFor) | 29 · Check that the primary comes back | 🆕 New lab |
Same-backend retries and the per-request retry limit (retryAttempts, maxRetryAttempts) | 29 · How it works | 🆕 New lab |
| Failure handling in the Create alias wizard and the alias editor | 29 · Add a failure rule | 🆕 New lab |
| Priority alone passes backend errors to the client; add failure rules to fail over | 22 · How it works | ✏️ Updated |
Routing attributes and metrics in telemetry (azure.ai_gateway.routing.*) | 13 · Read the routing attributes | ✏️ Updated |
November 9, 2026
Jev decisions and per-request model selection
Publish TypeSafe Jev decision models behind an alias, and let callers override auto model selection on individual requests. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Publish TypeSafe Jev models behind a gateway alias with a Custom provider | 31 · Publish Jev behind an alias | 🆕 New lab |
| Evaluate yes/no questions, categorical choices, and scoring rubrics with native System One requests over uncompressed JSON | 31 · Send a decision request | 🆕 New lab |
| Limit decision requests per caller | 31 · Limit decision requests per caller | 🆕 New lab |
| Allow callers to override the objective or narrow the model pool for one request (preview) | 28 · Let callers choose per request | 🆕 New lab |
| Filter auto model selection for function-tool or request-parameter compatibility | Explained in 28 · Go further | ➖ Not covered |
| Priority and weighted routing, and decision telemetry, for Jev models | Explained in 31 · Go further and 13 · Read the routing attributes | ➖ Not covered |
November 6, 2026
Auto model selection, session affinity, Agent Hooks, and on-behalf-of authentication
Combine prompt-aware model selection with gateway routing, and add delegated authentication and custom governance to MCP backends. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Choose a model for each prompt with auto model selection and a Balanced, Cost, or Quality objective (preview) | 28 · Let the gateway choose the model for each request | 🆕 New lab |
See which model served a request with x-ms-served-model | 28 · Send a simple and a hard prompt | 🆕 New lab |
| Route an alias's requests by request headers or the caller with CEL backend conditions | 27 · Route research requests to the larger model | 🆕 New lab |
| Keep a conversation on one backend model with session affinity | 27 · Keep each session on one backend | 🆕 New lab |
| Write and check CEL expressions in the portal | 27 · Create the chat alias with a condition | 🆕 New lab |
| Understand the routing order: conditions, priority, auto selection, session affinity, and weights | 22 · How it works and 27 · Go further | ✏️ Updated |
| Stateful API continuations stay pinned to their owning backend | 23 · How it works | ✏️ Updated |
Ask your decision service before an MCP tool runs, and enforce its allow or deny verdict (Agent Hooks, preview) | 30 · Call a blocked tool | 🆕 New lab |
| Screen tool results before they reach the agent with an after-tool-call hook | 30 · Get a result that contains a secret | 🆕 New lab |
| Choose fail closed or fail open when the decision service is unavailable | 30 · See the policy fail closed | 🆕 New lab |
| Grant management, read, and runtime access separately with the control-plane roles | 11 · Grant control-plane access | ✏️ Updated |
| Assign runtime roles on all assets, a model alias, or an MCP server | 11 · Grant a runtime role | ✏️ Updated |
Mark failed model streams and MCP tool calls as Error in telemetry, even when the response is 200 | 13 · Find failed operations | ✏️ Updated |
| Count failed operations per caller in a chargeback report | 14 · Build a chargeback report | ✏️ Updated |
| Call MCP backends as the signed-in user with on-behalf-of token exchange | 33 · Connect Microsoft 365 Copilot to MCP tools as the signed-in user, explained in 12 · Call MCP backends as the signed-in user | 🆕 New lab |
| Edit virtual network settings after gateway creation: add, change, or remove the integration subnet | 18 · Remove and add back private routing and 01 · Review the template | ✏️ Updated |
Send runtime keys as Authorization: Bearer <key> | 10 · Send the key in either header and 02 · How it works | ✏️ Updated |
Fixed: stream Anthropic Messages without stream_options failures | 25 · Ask a quick question | ✅ Covered |
| The new routing options, Agent Hooks, control-plane roles, and failure telemetry in the big picture | 00 · How Azure AI Gateway works | ✏️ Updated |
October 28, 2026
Embedding models
Import embedding models, publish stable aliases, and generate vectors in the portal. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Import embedding deployments from Foundry and publish them behind stable aliases | 26 · Publish embedding models for search and RAG | 🆕 New lab |
| Generate vectors with text batches and backend-supported dimensions | 26 · Generate embeddings | 🆕 New lab |
| Test embeddings in the Catalog playground and Continue in code | 26 · Test in the playground | 🆕 New lab |
| Apply token limits and usage telemetry to embedding models | 26 · Cap ingestion with a token limit | 🆕 New lab |
| Route an embeddings alias across deployments by priority and weight | Explained in 26 · Go further, practiced with chat models in 22 · Route an alias across models by priority and weight | ➖ Not covered |
October 16, 2026
Priority and weighted model routing, MCP 2026-07-28
Keep a stable alias while you prefer purchased capacity, balance traffic across deployments, or test a new model version, and manage MCP servers that use the 2026-07-28 MCP specification. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Route an alias's traffic across backend models by priority and weight | 22 · Route an alias across models by priority and weight | 🆕 New lab |
| Canary a new model version or run an A/B test with a controlled split | 22 · Split traffic 80/20 between two models | 🆕 New lab |
| Prefer purchased capacity with priorities | 22 · Prefer one model by priority | 🆕 New lab |
| Create and edit aliases with the Create alias wizard | 03 · Publish models from any provider behind stable aliases | ✏️ Updated |
| Choose the MCP protocol versions a server supports, including 2026-07-28 | 04 · Choose supported MCP versions | ✏️ Updated |
| See supported MCP versions on the server page and in Catalog | 04 · Check the negotiated version | ✏️ Updated |
October 8, 2026
Agent 365 integration and tool threat protection
Discover and govern AI Gateway MCP servers and tools from Microsoft Agent 365, and protect tool calls with Microsoft Defender and Microsoft Purview. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Discover AI Gateway MCP servers in the Agent 365 inventory | 20 · Connect Agent 365 to your gateways | ✅ Covered |
| Block or unblock MCP servers and individual tools from Agent 365 | 20 · Block a tool and 20 · Block the whole server | ✅ Covered |
| Evaluate tool calls before and after invocation with Defender and Purview | 19 · Add tool threat protection and 19 · Understand verdicts and failures | ✅ Covered |
October 1, 2026
Cost limit overrides for Microsoft Entra identities
Set different cost limit amounts for Microsoft Entra callers by identity or by an exact combination of validated token claims. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| A higher or lower budget for a selected identity or matching token claims | 07 · Give a Microsoft Entra identity its own budget | ✅ Covered |
| Separate budgets per caller by tenant ID and object ID | 07 · Enforce cost budgets with per-caller overrides | ✅ Covered |
| The first matching override wins | 07 · How it works | ✅ Covered |
September 24, 2026
Model aliases, Entra access, and Chat Completions streaming usage
Aliases become the stable model contract, with Microsoft Entra authentication, gateway-scoped policies, and complete usage for streamed Chat Completions responses. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| Callable aliases, several aliases for one backend model, and retargeting | 03 · Publish models from any provider behind stable aliases | ✅ Covered |
| Grant model and MCP access to Microsoft Entra users, groups, apps, and managed identities | 11 · Go keyless with Microsoft Entra ID and RBAC | ✅ Covered |
| Export selected JWT claims to OTLP logs, traces, and metrics | 14 · Choose the claims to export | ✅ Covered |
| Apply one policy to every current and future asset, including a gateway-wide cost limit | 09 · Set gateway-wide policy baselines | ✅ Covered |
Usage for streamed Chat Completions responses without include_usage | 23 · Stream a burst without asking for usage | 🆕 New lab |
September 10, 2026
Per-key budget overrides and MCP tool filtering and blocking
Assign higher or lower model budgets to selected API keys, and control which MCP tools agents can discover and invoke. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| A different budget for selected API keys | 07 · Give an API key a bigger budget | ✅ Covered |
| Expose all, selected, or pattern-matched tools | 05 · Publish only approved tools | ✅ Covered |
| Block individual tools or a whole MCP server | 05 · Take the whole server offline | ✅ Covered |
| Redesigned Home and first-run gateway setup | 01 · Provision your AI Gateway | ✅ Covered |
August 28, 2026
OpenTelemetry observability and model budgets
Export logs, traces, token metrics, and estimated cost metrics through OTLP, and enforce model spend with cost limits. Read the product release notes.
| In the product | Practice it in the labs | Status |
|---|---|---|
| The built-in Application Insights dashboard and request traces | 13 · Monitor usage, cost, and traces | ✅ Covered |
| OTLP export of logs, traces, and token and cost metrics | 14 · Attribute tokens and cost with OpenTelemetry | ✅ Covered |
| Cost limits with hourly to yearly budget windows | 07 · Enforce cost budgets with per-caller overrides | ✅ Covered |
| Import Microsoft Foundry models and onboard custom providers | 02 · Publish a Foundry model and 03 · Add Anthropic as a custom provider | ✅ Covered |
How these notes work
Each section matches one entry in the product release notes, with the same date and title, and maps every product change to the lab where you can practice it.
This file, labs/RELEASE-NOTES.md, is the source of truth: the labs site renders it at
/release-notes/. To check for product releases that the
notes don't cover yet, run npm run releases in labs/. See Authoring.