The Azure Updates entry itself did not render when the page was fetched. The details below come from Microsoft Learn and the extension's GitHub repository, which is Microsoft-owned.
What is retiring
The extension's repository now logs the April 20, 2027 retirement notice at Warning level during extension initialization. The project added OpenAI LLM bindings (GPT-3.5-turbo, GPT-4, o-series) to Azure Functions. Microsoft Learn describes it as a set of triggers and bindings for using Azure OpenAI in Foundry Models from function code.
The documented bindings are:
- Text completion input binding
- Assistant trigger, assistant create output binding, assistant post input binding and assistant query input binding
- Embeddings input binding
- Embeddings store output binding, for writing to a vector database
- Semantic search input binding, for reading from a vector database
Microsoft's documentation doesn't say that other Functions triggers, bindings or apps are affected. Treat the scope as projects that use these OpenAI bindings.
The earlier cutoff: bundle 4.60.0
The retirement date isn't the first thing that can break an app. Per Microsoft Learn, preview extension bundle version 4.60.0 and later don't include the extension. An app that floats on the preview bundle could therefore lose the bindings when it picks up a newer bundle, before April 2027.
Microsoft's stopgap for existing apps is to set the bundle version range in host.json:
{
"version": "2.0",
"extensionBundle": {
"id": "Microsoft.Azure.Functions.ExtensionBundle.Preview",
"version": "[4.*, 4.60.0)"
}
}
The range keeps the app on preview bundles earlier than 4.60.0. The repository's own pull request pins all 20 bundle-based sample host.json files to the same range.
What the workarounds don't do
Microsoft is blunt on this point. Pinning an older bundle doesn't extend support beyond April 20, 2027. Neither does continuing to use the extension's NuGet packages directly in .NET apps. Both options only delay breakage. Neither gets you support after the date.
What Microsoft recommends instead
Microsoft says to move to the OpenAI, Microsoft Foundry, or Azure AI Search SDKs directly, depending on your scenario. It doesn't publish a one-to-one mapping from each binding to an SDK call, and it doesn't say code will migrate automatically. You'll need to work out that mapping yourself.
A practical migration checklist
- Inventory usage. Search your Function projects for the extension's packages and bindings. For .NET, look for
Microsoft.Azure.Functions.Worker.Extensions.OpenAI(isolated worker) orMicrosoft.Azure.WebJobs.Extensions.OpenAI(in-process). Also look for the vector-store add-ons for Azure AI Search, Cosmos DB and Azure Data Explorer. This step is my inference from Microsoft's feature list, not a Microsoft procedure. - Check
host.json. If you use the preview bundle, confirm your version range. Apply the[4.*, 4.60.0)pin now if you can't migrate immediately. - Map each binding to an SDK. Completions and assistant calls likely map to the OpenAI or Foundry SDKs. Vector search and embeddings storage are the likely fit for the Azure AI Search SDK. This mapping is my reading of Microsoft's guidance. Microsoft doesn't prescribe it.
- Redo connection handling. The extension uses an
AIConnectionNameproperty, plusAZURE_OPENAI_ENDPOINTandAZURE_OPENAI_KEYapp settings. Microsoft says that when you move to direct SDK calls, you should follow the authentication guidance for the SDK you choose, not the extension's binding settings. Microsoft's guidance for the extension prefers managed identity over shared keys. That is also a sensible default for the replacement code. - Test and cut over before April 20, 2027. Microsoft Learn says Go isn't supported for this feature. The repository has samples for .NET isolated, Java, Node.js, PowerShell and Python.
Analysis
The extension was always a convenience layer, and the Learn page tells you to install it with the --prerelease flag. It was never a settled part of the Functions platform. Teams that built production workloads on it are now paying for that choice.
The upside is that the migration is mostly code you'd own anyway. Calling an SDK directly gives you control over model parameters, retries and authentication. The cost is the glue code the bindings used to supply, especially for assistant state and vector-store reads and writes.
Microsoft has about five months of notice here, counting from this announcement to the retirement date. That is enough time if you start the inventory now, and tight if you wait. Because the bundle cutoff can bite earlier than the retirement date, check your host.json this week.
References
- Retirement: Support for Azure Functions OpenAI Extension ends on April 20, 2027 Azure Updates · 2026-11-20T23:16:31Z
- GitHub - Azure/azure-functions-openai-extension: An extension that adds support for Azure OpenAI/ OpenAI bindings in Azure Functions for LLM (GPT-3.5-Turbo, GPT-4, etc) · GitHub github.com
- Add OpenAI extension retirement warning and pin sample bundles by manvkaur · Pull Request #228 · Azure/azure-functions-openai-extension · GitHub github.com