@ai-sdk/openai 4.0.68 → 4.0.70

package/docs/03-openai.mdx

@@ -3397,3 +3397,48 @@ const result = await generateSpeech({
 | `tts-1`           | <Check />    |
 | `tts-1-hd`        | <Check />    |
 | `gpt-4o-mini-tts` | <Check />    |
+
+## Evaluation Models
+
+Create an experimental evaluation model with `openai.evaluationModel(modelId)`.
+It uses the Responses API with structured output to answer Choice, Score, and
+Boolean questions. Choose a model that supports structured output, such as `gpt-5.6-luna`.
+
+```ts
+import { openai } from '@ai-sdk/openai';
+import { experimental_evaluate } from 'ai';
+
+const { answers } = await experimental_evaluate({
+  model: openai.evaluationModel('gpt-5.6-luna'),
+  state: 'I was charged twice.',
+  questions: {
+    requestsRefund: {
+      type: 'boolean',
+      instructions: 'Is the customer requesting money back?',
+    },
+    department: {
+      type: 'choice',
+      instructions: 'Which team should handle this?',
+      criteria: { billing: 'Charges and refunds', support: 'Other requests' },
+    },
+  },
+});
+```
+
+Scores are finite fractional positions on the ordered rubric. Choice labels are
+returned exactly as supplied. These are model judgments: no probability
+distributions are returned for Choice or Score. Boolean answers contain a prompted
+estimate of P(true), validated to be finite and in `[0, 1]`. These estimates are
+not guaranteed to be calibrated. Apply thresholds in application code, for example
+`answers.requestsRefund.probability >= 0.5`.
+
+The factory respects `createOpenAI` settings, including custom fetch and base URL.
+Pass Responses options through `providerOptions.openai`. Usage, warnings,
+response information, and provider metadata are preserved. Refusals, truncated
+output, or invalid answers fail the entire call. See [Evaluation](/docs/ai-sdk-core/evaluation).
+
+Evaluation models can also be accessed through `customProvider` aliases or
+`createProviderRegistry().evaluationModel('provider:model')`. Direct string IDs
+use Gateway by default, or an explicitly configured default provider with an
+`evaluationModel` method. See
+[model aliases and registries](/docs/ai-sdk-core/evaluation#model-aliases-and-registries).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/openai",
-  "version": "4.0.68",
+  "version": "4.0.70",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -35,8 +35,8 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/provider": "4.0.16",
-    "@ai-sdk/provider-utils": "5.0.42"
+    "@ai-sdk/provider": "4.0.17",
+    "@ai-sdk/provider-utils": "5.0.44"
   },
   "devDependencies": {
     "@ai-sdk/test-server": "2.0.1",

package/src/openai-provider.ts

@@ -1,5 +1,6 @@
 import type {
   Experimental_BatchV4 as BatchV4,
+  Experimental_EvaluationModelV4 as EvaluationModelV4,
   EmbeddingModelV4,
   FilesV4,
   ImageModelV4,
@@ -19,6 +20,7 @@ import {
   type FetchFunction,
   type WebSocketConstructor,
 } from '@ai-sdk/provider-utils';
+import { Experimental_EvaluationLanguageModel as EvaluationLanguageModel } from '@ai-sdk/provider-utils/experimental-evaluation';
 import { OpenAIChatLanguageModel } from './chat/openai-chat-language-model';
 import type { OpenAIChatModelId } from './chat/openai-chat-language-model-options';
 import { OpenAICompletionLanguageModel } from './completion/openai-completion-language-model';
@@ -48,6 +50,9 @@ import { VERSION } from './version';
 export interface OpenAIProvider extends ProviderV4 {
   (modelId: OpenAIResponsesModelId): LanguageModelV4;
 
+  /** Creates an experimental Choice/Score/Boolean evaluation model using the Responses API. */
+  evaluationModel(modelId: OpenAIResponsesModelId): EvaluationModelV4;
+
   /**
    * Creates an OpenAI model for text generation.
    */
@@ -347,6 +352,11 @@ export function createOpenAI(
   provider.chat = createChatModel;
   provider.completion = createCompletionModel;
   provider.responses = createResponsesModel;
+  provider.evaluationModel = (modelId: OpenAIResponsesModelId) =>
+    new EvaluationLanguageModel({
+      model: createResponsesModel(modelId),
+      provider: `${providerName}.evaluation`,
+    });
   provider.embedding = createEmbeddingModel;
   provider.embeddingModel = createEmbeddingModel;
   provider.textEmbedding = createEmbeddingModel;

package/src/responses/convert-to-openai-responses-input.ts

@@ -137,6 +137,32 @@ async function convertFunctionToolResultOutput({
               const imageDetail =
                 item.providerOptions?.[providerOptionsName]?.imageDetail;
 
+              if (item.data.type === 'reference') {
+                const fileId = resolveProviderReference({
+                  reference: item.data.reference,
+                  provider: providerOptionsName,
+                });
+
+                if (topLevel === 'image') {
+                  return {
+                    type: 'input_image' as const,
+                    file_id: fileId,
+                    detail: imageDetail,
+                    ...(promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint,
+                    }),
+                  };
+                }
+
+                return {
+                  type: 'input_file' as const,
+                  file_id: fileId,
+                  ...(promptCacheBreakpoint != null && {
+                    prompt_cache_breakpoint: promptCacheBreakpoint,
+                  }),
+                };
+              }
+
               if (item.data.type === 'data') {
                 const fullMediaType = resolveFullMediaType({ part: item });
                 if (topLevel === 'image') {
@@ -611,8 +637,7 @@ export async function convertToOpenAIResponsesInput({
               input.push({
                 ...(explicitMessageItemType && { type: 'message' as const }),
                 role: 'assistant',
-                content: [{ type: 'output_text', text: part.text }],
-                id,
+                content: part.text,
                 ...(phase != null && { phase }),
               });
 

package/src/responses/openai-responses-api.ts

@@ -266,8 +266,7 @@ export type OpenAIResponsesUserMessage = {
 export type OpenAIResponsesAssistantMessage = {
   type?: 'message';
   role: 'assistant';
-  content: Array<{ type: 'output_text'; text: string }>;
-  id?: string;
+  content: string;
   phase?: 'commentary' | 'final_answer' | null;
 };
 
@@ -298,6 +297,11 @@ export type OpenAIResponsesFunctionCallOutput = {
             image_url: string;
             prompt_cache_breakpoint?: { mode: 'explicit' };
           }
+        | {
+            type: 'input_image';
+            file_id: string;
+            prompt_cache_breakpoint?: { mode: 'explicit' };
+          }
         | {
             type: 'input_file';
             filename: string;
@@ -309,6 +313,11 @@ export type OpenAIResponsesFunctionCallOutput = {
             file_url: string;
             prompt_cache_breakpoint?: { mode: 'explicit' };
           }
+        | {
+            type: 'input_file';
+            file_id: string;
+            prompt_cache_breakpoint?: { mode: 'explicit' };
+          }
       >;
   caller?: OpenAIResponsesToolCaller;
 };