Prepare 2.9.0 release

joseharriaga · Feb 27, 2026 · Feb 27, 2026 · Feb 27, 2026 · Feb 27, 2026 · Feb 27, 2026
commit 680ea9af84e9a54ee9c907f9cb5e44ed6b5b75bc
@@ -1,5 +1,177 @@
 # Release History
 
+## 2.9.0 (2026-02-27)
+
+### Acknowledgments
+
+Thank you to our developer community members who helped to make the OpenAI client library better with their contributions to this release:
+
+- [@mokarchi](https://github.com/mokarchi)
+- [@alty-ir](https://github.com/alty-ir)
+- [@peibekwe](https://github.com/peibekwe)
+- [@crickman](https://github.com/crickman)
+- [@jpaarhuis](https://github.com/jpaarhuis)
+- [@salmanmkc](https://github.com/salmanmkc)
+
+### Features Added
+
+- OpenAI.Audio:
+  - Added support for `AudioClient` configuration and dependency injection via the new `AudioClientSettings` class.
+- OpenAI.Batch:
+  - Added support for `BatchClient` configuration and dependency injection via the new `BatchClientSettings` class.
+- OpenAI.Chat:
+  - Added support for `ChatClient` configuration and dependency injection via the new `ChatClientSettings` class.
+  - Added value of `None` to `ChatReasoningEffortLevel` enum to allow opting out of reasoning.
+- OpenAI.Containers:
+  - Added support for `ContainerClient` configuration and dependency injection via the new `ContainerClientSettings` class.
+- OpenAI.Conversations:
+  - Added support for `ConversationClient` configuration and dependency injection via the new `ConversationClientSettings` class.
+- OpenAI.Embeddings:
+  - Added support for `EmbeddingClient` configuration and dependency injection via the new `EmbeddingClientSettings` class.
+- OpenAI.Evals:
+  - Added support for `EvaluationClient` configuration and dependency injection via the new `EvaluationClientSettings` class.
+- OpenAI.Files:
+  - Added support for `OpenAIFileClient` configuration and dependency injection via the new `OpenAIFileClientSettings` class.
+  - Added support for retrieving files using pagination via new overloads of the `GetFiles` and `GetFilesAsync` methods in the `OpenAIFileClient`.
+- OpenAI.FineTuning:
+  - Added support for `FineTuningClient` configuration and dependency injection via the new `FineTuningClientSettings` class.
+- OpenAI.Graders:
+  - Added support for `GraderClient` configuration and dependency injection via the new `GraderClientSettings` class.
+- OpenAI.Images:
+  - Added support for `ImageClient` configuration and dependency injection via the new `ImageClientSettings` class.
+  - Added value of `HighQuality` to `GeneratedImageQuality` enum to allow generating higher quality images.
+  - Added `InputFidelity` property to `ImageEditOptions` to control how closely distinctive features from the input are preserved.
+  - Added `OutputCompressionFactor` property to `ImageEditOptions` to set the compression level (0-100%) for `jpeg` or `webp` outputs.
+  - Added `OutputFileFormat` property to `ImageEditOptions` to set the file format in which the generated images are returned.
+  - Added `Background` property to `GeneratedImageCollection`.
+  - Added `OutputFileFormat` property to `GeneratedImageCollection`.
+  - Added `Quality` property to `GeneratedImageCollection`.
+  - Added `Size` property to `GeneratedImageCollection`.
+- OpenAI.Models:
+  - Added support for `OpenAIModelClient` configuration and dependency injection via the new `OpenAIModelClientSettings` class.
+- OpenAI.Moderations:
+  - Added support for `ModerationClient` configuration and dependency injection via the new `ModerationClientSettings` class.
+  - Added support for classifying multi-modal inputs via new `ClassifyInputs` and `ClassifyInputsAsync` methods in the `ModerationClient`.
+    - Use the new `ModerationInputPart` class to create text or image inputs for classication.
+    - Use the new `ApplicableInputKinds` property of the `ModerationCategory` class to understand to which parts of the input the moderation results apply.
+- OpenAI.Realtime:
+  - Added support for the stable version of the Realtime REST API.
+  - Added support for `RealtimeClient` configuration and dependency injection via the new `RealtimeClientSettings` class.
+  - Added support for integrating with remote MCP servers.
+    - Users can add the new `RealtimeMcpTool` to the `Tools` property of their `RealtimeConversationSessionOptions` and configure it.
+      - Use the `AllowedTools` property to limit which of the server tools can be called by the model.
+      - Use the `ToolCallApprovalPolicy` property to specify which tools require an explicit approval before being called by the model.
+      - Use the `AuthorizationToken` property to specify the access token to used when authenticating with the remote MCP server.
+      - Use the `ConnectorId` property to use the OpenAI-maintained MCP wrappers for popular services like Microsoft Outlook or Dropbox.
+  - Added a new `RealtimeClientCommand` class that represents client events to be used with the new `SendCommand` and `SendCommandAsync` overloads in the `RealtimeSessionClient` and exposed all supported client events as derived types:
+    - `RealtimeClientCommandConversationItemCreate`
+    - `RealtimeClientCommandConversationItemDelete`
+    - `RealtimeClientCommandConversationItemRetrieve`
+    - `RealtimeClientCommandConversationItemTruncate`
+    - `RealtimeClientCommandInputAudioBufferAppend`
+    - `RealtimeClientCommandInputAudioBufferClear`
+    - `RealtimeClientCommandInputAudioBufferCommit`
+    - `RealtimeClientCommandOutputAudioBufferClear`
+    - `RealtimeClientCommandResponseCancel`
+    - `RealtimeClientCommandResponseCreate`
+    - `RealtimeClientCommandSessionUpdate`
+  - Added support for System.Client.Model's `JsonPatch`, which enables users to get and set additional JSON properties in response and request payloads.
+- OpenAI.Responses:
+  - Added support for `ResponsesClient` configuration and dependency injection via the new `ResponsesClientSettings` class.
+  - Added support for compacting conversations via the new `CompactResponse` and `CompactResponseAsync` protocol methods in the `ResponsesClient`.
+  - Added support for returning the number of input tokens that a Response would consume via the new `GetInputTokenCount` and `GetInputTokenCountAsync` protocol methods in the `ResponsesClient`.
+  - Added value of `None` to `ResponseReasoningEffortLevel` enum to allow opting out of reasoning.
+  - Added `InputImageUri` property to `ResponseContentPart` to manage image inputs as URIs.
+  - Added implicit conversion operators to `McpToolCallApprovalPolicy` to facilitate its assignment.
+- OpenAI.VectorStores:
+  - Added support for `VectorStoreClient` configuration and dependency injection via the new `VectorStoreClientSettings` class.
+- OpenAI.Videos:
+  - Added support for `VideoClient` configuration and dependency injection via the new `VideoClientSettings` class.
+
+### Bugs Fixed
+
+- OpenAI.Chat:
+  - Fixed a concurrency issue corrupting the state of the `ChatCompletionOptions` argument when the same instance was passed into concurrent calls of `CompleteChat`/`CompleteChatAsync` and/or `CompleteChatStreaming`/`CompleteChatStreamingAsync`.
+- OpenAI.Realtime:
+  - Fixed an issue causing permanent state corruption when sending audio was cancelled in `RealtimeSessionClient`.
+  - Fixed an issue with `RealtimeSessionClient` always requiring an API key.
+- OpenAI.Responses:
+  - Fixed a concurrency issue corrupting the state of the `CreateResponseOptions` argument when the same instance was passed into concurrent calls of `CreateResponse`/`CreateResponseAsync` and/or `CreateResponseStreaming`/`CreateResponseStreamingAsync`.
+
+### Other Changes
+
+- Updated the `System.ClientModel` dependency to version 1.9.0. For more information, see the [System.ClientModel Changelog](https://github.com/Azure/azure-sdk-for-net/blob/System.ClientModel_1.9.0/sdk/core/System.ClientModel/CHANGELOG.md)
+
+### Breaking Changes in Preview APIs
+
+- OpenAI.Images:
+  - Changed the type of the `InputTokenCount`, `OutputTokenCount`, and `TotalTokenCount` properties of `ImageTokenUsage` from `int` to `long`.
+- OpenAI.Realtime:
+  - This release migrates the implementation of the Realtime API from the beta REST API to the GA REST API. As such, it contains numerous breaking changes, including the following:
+    - Renamed `RealtimeSession` to `RealtimeSessionClient`.
+    - Renamed `ConversationTool` to `RealtimeTool` and exposed all supported tools as derived types:
+      - `RealtimeFunctionTool`
+      - `RealtimeMcpTool`
+    - Changed `RealtimeItem` into a class hierarchy and exposed all supported item types as derived types:
+      - `RealtimeMessageItem`
+      - `RealtimeFunctionCallItem`
+      - `RealtimeFunctionCallOutputItem`
+      - `RealtimeMcpToolCallItem`
+      - `RealtimeMcpToolDefinitionListItem`
+      - `RealtimeMcpToolCallApprovalRequestItem`
+      - `RealtimeMcpToolCallApprovalResponseItem`
+    - Renamed `RealtimeUpdate` to `RealtimeServerUpdate` and exposed all supported server events as derived types:
+      - `RealtimeServerUpdateConversationCreated`
+      - `RealtimeServerUpdateConversationItemAdded`
+      - `RealtimeServerUpdateConversationItemCreated`
+      - `RealtimeServerUpdateConversationItemDeleted`
+      - `RealtimeServerUpdateConversationItemDone`
+      - `RealtimeServerUpdateConversationItemInputAudioTranscriptionCompleted`
+      - `RealtimeServerUpdateConversationItemInputAudioTranscriptionDelta`
+      - `RealtimeServerUpdateConversationItemInputAudioTranscriptionFailed`
+      - `RealtimeServerUpdateConversationItemInputAudioTranscriptionSegment`
+      - `RealtimeServerUpdateConversationItemRetrieved`
+      - `RealtimeServerUpdateConversationItemTruncated`
+      - `RealtimeServerUpdateError`
+      - `RealtimeServerUpdateInputAudioBufferCleared`
+      - `RealtimeServerUpdateInputAudioBufferCommitted`
+      - `RealtimeServerUpdateInputAudioBufferDtmfEventReceived`
+      - `RealtimeServerUpdateInputAudioBufferSpeechStarted`
+      - `RealtimeServerUpdateInputAudioBufferSpeechStopped`
+      - `RealtimeServerUpdateInputAudioBufferTimeoutTriggered`
+      - `RealtimeServerUpdateMcpListToolsCompleted`
+      - `RealtimeServerUpdateMcpListToolsFailed`
+      - `RealtimeServerUpdateMcpListToolsInProgress`
+      - `RealtimeServerUpdateOutputAudioBufferCleared`
+      - `RealtimeServerUpdateOutputAudioBufferStarted`
+      - `RealtimeServerUpdateOutputAudioBufferStopped`
+      - `RealtimeServerUpdateRateLimitsUpdated`
+      - `RealtimeServerUpdateResponseContentPartAdded`
+      - `RealtimeServerUpdateResponseContentPartDone`
+      - `RealtimeServerUpdateResponseCreated`
+      - `RealtimeServerUpdateResponseDone`
+      - `RealtimeServerUpdateResponseFunctionCallArgumentsDelta`
+      - `RealtimeServerUpdateResponseFunctionCallArgumentsDone`
+      - `RealtimeServerUpdateResponseMcpCallArgumentsDelta`
+      - `RealtimeServerUpdateResponseMcpCallArgumentsDone`
+      - `RealtimeServerUpdateResponseMcpCallCompleted`
+      - `RealtimeServerUpdateResponseMcpCallFailed`
+      - `RealtimeServerUpdateResponseMcpCallInProgress`
+      - `RealtimeServerUpdateResponseOutputAudioDelta`
+      - `RealtimeServerUpdateResponseOutputAudioDone`
+      - `RealtimeServerUpdateResponseOutputAudioTranscriptDelta`
+      - `RealtimeServerUpdateResponseOutputAudioTranscriptDone`
+      - `RealtimeServerUpdateResponseOutputItemAdded`
+      - `RealtimeServerUpdateResponseOutputItemDone`
+      - `RealtimeServerUpdateResponseOutputTextDelta`
+      - `RealtimeServerUpdateResponseOutputTextDone`
+      - `RealtimeServerUpdateSessionCreated`
+      - `RealtimeServerUpdateSessionUpdated`
+  - We recommend checking out our [examples](https://github.com/openai/openai-dotnet/tree/main/examples/Realtime) and tests as well as the [migration guide](https://developers.openai.com/api/docs/guides/realtime/#beta-to-ga-migration) for tips on how to use the new API surface.
+- OpenAI.Responses:
+  - Removed the `Model` property of the `ResponsesClient` in favor of specifying the model on a per-call basis when calling `CreateResponse`/`CreateResponseAsync` or `CreateResponseStreaming`/`CreateResponseStreamingAsync`. To do this, set the `Model` property of the `CreateResponseOptions` parameter. With this change, users no longer have to provide a model unless it is actually needed.
+  - Changed the type of the `Instructions` property of `ResponseResult` from `string` to `IList<ResponseItem>` to handle multi-item instructions.
+
 ## 2.8.0 (2025-12-11)
 
 ### Features Added

@@ -47,8 +47,8 @@
 
   <!-- Versioning -->
   <PropertyGroup>
-    <VersionPrefix>2.8.0</VersionPrefix>
-    <VersionSuffix></VersionSuffix>
+    <VersionPrefix>2.9.0</VersionPrefix>
+    <VersionSuffix>alpha.20260226.1</VersionSuffix>
     <InformationalVersion>$(Version)</InformationalVersion>
   </PropertyGroup>
 

@@ -69,6 +69,9 @@ function Invoke-APICompat {
         [Parameter(Mandatory = $true)]
         [string]$ProjectPath,
 
+        [Parameter(Mandatory = $true)]
+        [string]$DirectoryBuildPropsPath,
+
         [Parameter(Mandatory = $true)]
         [string]$ReleasePath,
 
@@ -89,9 +92,9 @@ function Invoke-APICompat {
         Invoke-DotNetPack -ProjectPath $ProjectPath
 
         # Extract the values of VersionPrefix and VersionSuffix from the .csproj XML file.
-        $xml = [xml](Get-Content $ProjectPath)
-        $versionPrefix = $($xml.Project.PropertyGroup[0].VersionPrefix)
-        $versionSuffix = $($xml.Project.PropertyGroup[0].VersionSuffix)
+        $xml = [xml](Get-Content $DirectoryBuildPropsPath)
+        $versionPrefix = $($xml.Project.PropertyGroup[3].VersionPrefix)
+        $versionSuffix = $($xml.Project.PropertyGroup[3].VersionSuffix)
         $currentVersion = [string]::IsNullOrEmpty($versionSuffix) ? "$($versionPrefix)" : "$($versionPrefix)-$($versionSuffix)"
 
         $currentNuGetPackagePath = Join-Path $ReleasePath "$($PackageName).$($currentVersion).nupkg"
@@ -160,6 +163,7 @@ function Invoke-APICompat {
 
 $repoRootPath = Join-Path $PSScriptRoot .. -Resolve
 $projectPath = Join-Path $repoRootPath "src\OpenAI.csproj"
+$buildPropsPath = Join-Path $repoRootPath "Directory.Build.props"
 $releasePath = Join-Path $repoRootPath "src\bin\Release"
 
 $experimentalNamespaces = @(
@@ -175,7 +179,8 @@ $experimentalNamespaces = @(
 )
 
 Invoke-APICompat -ProjectPath $projectPath `
+    -DirectoryBuildPropsPath $buildPropsPath `
     -ReleasePath $releasePath `
     -PackageName "OpenAI" `
-    -BaselineVersion "2.7.0" `
+    -BaselineVersion "2.8.0" `
     -IgnoredNamespaces $experimentalNamespaces
@@ -412,15 +412,12 @@ public async Task BasicRunStepFunctionalityWorks()
         Assert.That(details?.CreatedMessageId, Is.Not.Null.And.Not.Empty);
 
         details = secondStep.Details;
-        Assert.Multiple(() =>
-        {
-            Assert.That(details?.ToolCalls.Count, Is.GreaterThan(0));
-            Assert.That(details.ToolCalls[0].Kind, Is.EqualTo(RunStepToolCallKind.CodeInterpreter));
-            Assert.That(details.ToolCalls[0].Id, Is.Not.Null.And.Not.Empty);
-            Assert.That(details.ToolCalls[0].CodeInterpreterInput, Is.Not.Null.And.Not.Empty);
-            Assert.That(details.ToolCalls[0].CodeInterpreterOutputs?.Count, Is.GreaterThan(0));
-            Assert.That(details.ToolCalls[0].CodeInterpreterOutputs[0].ImageFileId, Is.Not.Null.And.Not.Empty);
-        });
+        Assert.That(details?.ToolCalls.Count, Is.GreaterThan(0));
+        Assert.That(details.ToolCalls[0].Kind, Is.EqualTo(RunStepToolCallKind.CodeInterpreter));
+        Assert.That(details.ToolCalls[0].Id, Is.Not.Null.And.Not.Empty);
+        Assert.That(details.ToolCalls[0].CodeInterpreterInput, Is.Not.Null.And.Not.Empty);
+        Assert.That(details.ToolCalls[0].CodeInterpreterOutputs?.Count, Is.GreaterThan(0));
+        Assert.That(details.ToolCalls[0].CodeInterpreterOutputs[0].ImageFileId, Is.Not.Null.And.Not.Empty);
     }
 
     [RecordedTest]

@@ -303,7 +303,7 @@ public async Task StreamingTranscriptionWorks(AudioSourceKind audioSourceKind)
             }
         }
 
-        Assert.That(deltaBuilder.ToString().ToLower(), Does.Contain("hello world"));
+        Assert.That(deltaBuilder.ToString().ToLower(), Does.Contain("this is a test"));
 
         inputStream?.Dispose();
     }

@@ -1,4 +1,5 @@
 using NUnit.Framework;
+using OpenAI.Realtime;
 using System;
 using System.ClientModel;
 using System.ClientModel.Primitives;
@@ -24,7 +25,8 @@ public static IEnumerable<Type> GetClientTypes()
                 type.IsClass && 
                 !type.IsAbstract &&
                 type.Name.EndsWith("Client", StringComparison.Ordinal) &&
-                HasEndpointField(type))
+                HasEndpointField(type) &&
+                type != typeof(RealtimeSessionClient))
             .OrderBy(type => type.Name); // For consistent ordering in test results
     }
 

@@ -25,7 +25,7 @@ public async Task CreateUploadWorks(bool useTopLevelClient)
         // are setting the internal Uploads client correctly.
 
         OpenAIFileClient fileClient = useTopLevelClient
-            ? TestEnvironment.GetTestClient<OpenAIClient>().GetOpenAIFileClient()
+            ? GetProxiedOpenAIClient<OpenAIClient>().GetOpenAIFileClient()
             : GetProxiedOpenAIClient<OpenAIFileClient>();
         BinaryContent content = BinaryContent.Create(BinaryData.FromObjectAsJson(new
         {
-Original file line number
+Diff line change
@@ Expand Up @@
                 }
             }
-            Assert.That(deltaBuilder.ToString().ToLower(), Does.Contain("hello world"));
+            Assert.That(deltaBuilder.ToString().ToLower(), Does.Contain("this is a test"));
             inputStream?.Dispose();
         }
@@ Expand Down @@