Microsoft.Extensions.AI.OpenAI 10.10.1 Fixes the OpenAI 2.14 TypeLoadException

Microsoft.Extensions.AI.OpenAI 10.10.1 hit NuGet on September 25, 2026, and the release notes are one line: “Upgrade OpenAI SDK to 2.14.0”. That line hides a runtime crash. If your app references Microsoft.Extensions.AI.OpenAI 10.10.0 and anything in the graph pulled OpenAI up to 2.14.0 (a Dependabot bump, a direct reference, another package), every Responses API call that carries a tool has been dying with a TypeLoadException since September 15.

A renamed experimental type, bound at JIT time

OpenAI 2.14.0 shipped on September 15. Among its changes, the experimental struct OpenAI.Responses.GlobalMcpToolCallApprovalPolicy became DefaultMcpToolCallApprovalPolicy, and the GlobalPolicy property on McpToolCallApprovalPolicy became DefaultPolicy. Experimental APIs (OPENAI001) are allowed to break, but Microsoft.Extensions.AI.OpenAI 10.10.0 was compiled against the old name inside OpenAIResponsesChatClient.ToResponseTool, the method that maps every AITool to a Responses tool.

The 10.10.0 nuspec declares OpenAI with a minimum of 2.13.0, so NuGet happily resolves 2.14.0 when something else asks for it. Nothing fails at build time. The first call through a tool-carrying ChatOptions fails when the JIT compiles ToResponseTool:

System.TypeLoadException: Could not load type 'OpenAI.Responses.GlobalMcpToolCallApprovalPolicy'
from assembly 'OpenAI, Version=2.14.0.0, Culture=neutral, PublicKeyToken=b4187f3e65366280'.
   at Microsoft.Extensions.AI.OpenAIResponsesChatClient.ToResponseTool(AITool tool, ChatOptions options, ToolSearchLookup toolSearchLookup)
   at Microsoft.Extensions.AI.OpenAIResponsesChatClient.AsCreateResponseOptions(...)
   at Microsoft.Extensions.AI.OpenAIResponsesChatClient.GetStreamingResponseAsync(...)

It does not matter whether you use MCP. The method references the type, so a plain AIFunction triggers it too. Issue #7760 reported it with Microsoft.Agents.AI.OpenAI 1.21.0 on .NET 10.

Why staying on OpenAI 2.13.0 was not a clean workaround

Pinning OpenAI back to 2.13.0 avoids the crash, but 2.13.0 has its own bug: ReasoningResponseItem deserialization calls ToReasoningStatus() on a JSON null. Third-party OpenAI-compatible backends that serialize a reasoning item as "status": null instead of omitting it tear down the whole SSE stream with ArgumentOutOfRangeException: Unknown ReasoningStatus value. OpenAI 2.14.0 added the null guard. So for a week you had to pick which bug you wanted.

PR #7761 resolves both: it bumps to OpenAI 2.14.0, maps HostedMcpServerToolAlwaysRequireApprovalMode and HostedMcpServerToolNeverRequireApprovalMode onto DefaultMcpToolCallApprovalPolicy, and adds a streaming regression test for the null status case.

The upgrade

Move both packages together. Microsoft.Extensions.AI.OpenAI 10.10.1 now requires OpenAI 2.14.0, so if you pinned 2.13.0 as a workaround, drop that pin or you will get an NU1605 downgrade error:

<ItemGroup>
  <PackageReference Include="Microsoft.Extensions.AI.OpenAI" Version="10.10.1" />
  <PackageReference Include="OpenAI" Version="2.14.0" />
</ItemGroup>

Then check what actually resolved, since agent frameworks and SDK wrappers often bring these in transitively:

dotnet list package --include-transitive | grep -E "OpenAI|Extensions.AI"

If you touch the MCP approval types directly in your own code, the rename applies to you as well:

#pragma warning disable OPENAI001
var policy = new McpToolCallApprovalPolicy(DefaultMcpToolCallApprovalPolicy.NeverRequireApproval);
var mode = policy.DefaultPolicy; // was policy.GlobalPolicy in OpenAI 2.13.0
#pragma warning restore OPENAI001

The broader lesson: a library’s minimum-version dependency on a package with experimental APIs is effectively an open range for breaking changes. If you run a Responses-based agent in production, a smoke test that sends one tool-carrying request after every dependency bump would have caught this in CI instead of at runtime. For the rest of what changed in this release line, see the earlier write-up on Microsoft.Extensions.AI 10.10 failing unscored evaluation metrics.