@ai-sdk/openai 3.0.137 → 3.0.139

package/docs/03-openai.mdx

@@ -1636,6 +1636,47 @@ const result = await generateText({
 });
 ```
 
+#### Refusals
+
+The Responses API can answer a request with a `refusal` content part instead of
+`output_text`. The provider surfaces the refusal explanation as text, and
+reports a `content-filter` finish reason in both `generateText` and
+`streamText`. When the same response also contains a client-side function call,
+the finish reason is `tool-calls` instead, so the tools are still executed; the
+refusal text remains part of the content.
+
+Refusal content parts mainly appear with structured outputs (`Output.object` or
+a JSON schema passed to the model). With structured output, a refusal means no
+object can be generated: `generateText` throws `NoObjectGeneratedError` with the
+refusal text on `error.text` and `error.finishReason` set to `'content-filter'`.
+
+```ts
+import { NoObjectGeneratedError, Output, generateText } from 'ai';
+import { openai } from '@ai-sdk/openai';
+import { z } from 'zod';
+
+try {
+  await generateText({
+    model: openai.responses('gpt-4o'),
+    output: Output.object({
+      schema: z.object({ steps: z.array(z.string()) }),
+    }),
+    prompt: '...',
+  });
+} catch (error) {
+  if (NoObjectGeneratedError.isInstance(error)) {
+    // the model's refusal explanation
+    console.log(error.text);
+    // 'content-filter'
+    console.log(error.finishReason);
+  }
+}
+```
+
+Without structured output, the model may still decline in plain text. Such a
+response arrives as ordinary text with a `stop` finish reason, so check the text
+in your application if you need to detect declines.
+
 #### Typed providerMetadata in Text Parts
 
 When using the OpenAI Responses API, the SDK attaches OpenAI-specific metadata to output parts via `providerMetadata`.
@@ -2113,6 +2154,33 @@ You need to change Zod `.nullish()` and `.optional()` to `.nullable()`.
 
 </Note>
 
+#### Refusals
+
+Refusals mainly happen when you request structured outputs. When OpenAI declines
+a request, the provider surfaces the refusal explanation as text and reports a
+`content-filter` finish reason in both `generateText` and `streamText`:
+
+```ts
+const { text, finishReason } = await generateText({
+  model: openai.chat('gpt-4o'),
+  prompt: '...',
+});
+
+if (finishReason === 'content-filter') {
+  // `text` contains the model's refusal explanation
+}
+```
+
+When you request structured output, a refusal means that no object can be
+generated: `generateText` throws `NoObjectGeneratedError` with the refusal text
+on `error.text`.
+
+When you request logprobs, `providerMetadata.openai.logprobs` contains the
+logprobs of all emitted tokens: the content logprobs first, followed by the
+refusal logprobs. This is the same for `generateText` and `streamText`; with
+`streamText`, the logprobs of all streamed tokens are reported when the stream
+finishes.
+
 #### Logprobs
 
 OpenAI provides logprobs information for completion/chat models.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/openai",
-  "version": "3.0.137",
+  "version": "3.0.139",
   "license": "Apache-2.0",
   "sideEffects": false,
   "main": "./dist/index.js",
@@ -37,7 +37,7 @@
   },
   "dependencies": {
     "@ai-sdk/provider": "3.0.19",
-    "@ai-sdk/provider-utils": "4.0.64"
+    "@ai-sdk/provider-utils": "4.0.65"
   },
   "devDependencies": {
     "@types/node": "20.17.24",

package/src/chat/map-openai-finish-reason.ts

@@ -1,11 +1,18 @@
 import type { LanguageModelV3FinishReason } from '@ai-sdk/provider';
 
-export function mapOpenAIFinishReason(
-  finishReason: string | null | undefined,
-): LanguageModelV3FinishReason['unified'] {
+export function mapOpenAIFinishReason({
+  finishReason,
+  hasRefusal = false,
+}: {
+  finishReason: string | null | undefined;
+  // flag that checks if the model refused to answer:
+  hasRefusal?: boolean;
+}): LanguageModelV3FinishReason['unified'] {
   switch (finishReason) {
     case 'stop':
-      return 'stop';
+      // OpenAI reports a completed response for refusals, so the refusal is
+      // the reason the response ended:
+      return hasRefusal ? 'content-filter' : 'stop';
     case 'length':
       return 'length';
     case 'content_filter':

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

@@ -36,6 +36,8 @@ export const openaiChatResponseSchema = lazySchema(() =>
           message: z.object({
             role: z.literal('assistant').nullish(),
             content: z.string().nullish(),
+            // set instead of `content` when the model refuses to answer:
+            refusal: z.string().nullish(),
             audio: z
               .object({
                 transcript: z.string().nullish(),
@@ -84,6 +86,22 @@ export const openaiChatResponseSchema = lazySchema(() =>
                   }),
                 )
                 .nullish(),
+              // logprobs of the refusal tokens (set instead of `content` when
+              // the model refuses to answer):
+              refusal: z
+                .array(
+                  z.object({
+                    token: z.string(),
+                    logprob: z.number(),
+                    top_logprobs: z.array(
+                      z.object({
+                        token: z.string(),
+                        logprob: z.number(),
+                      }),
+                    ),
+                  }),
+                )
+                .nullish(),
             })
             .nullish(),
           finish_reason: z.string().nullish(),
@@ -128,6 +146,8 @@ export const openaiChatChunkSchema = lazySchema(() =>
               .object({
                 role: z.enum(['assistant']).nullish(),
                 content: z.string().nullish(),
+                // set instead of `content` when the model refuses to answer:
+                refusal: z.string().nullish(),
                 tool_calls: z
                   .array(
                     z.object({
@@ -172,6 +192,22 @@ export const openaiChatChunkSchema = lazySchema(() =>
                     }),
                   )
                   .nullish(),
+                // logprobs of the refusal tokens (set instead of `content` when
+                // the model refuses to answer):
+                refusal: z
+                  .array(
+                    z.object({
+                      token: z.string(),
+                      logprob: z.number(),
+                      top_logprobs: z.array(
+                        z.object({
+                          token: z.string(),
+                          logprob: z.number(),
+                        }),
+                      ),
+                    }),
+                  )
+                  .nullish(),
               })
               .nullish(),
             finish_reason: z.string().nullish(),

package/src/chat/openai-chat-language-model.ts

@@ -43,6 +43,15 @@ import {
 } from './openai-chat-options';
 import { prepareChatTools } from './openai-chat-prepare-tools';
 
+type OpenAIChatLogprob = NonNullable<
+  NonNullable<
+    Extract<
+      OpenAIChatChunk,
+      { choices: unknown }
+    >['choices'][number]['logprobs']
+  >['content']
+>[number];
+
 type OpenAIChatConfig = {
   provider: string;
   headers: () => Record<string, string | undefined>;
@@ -381,7 +390,10 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
 
     const content: Array<LanguageModelV3Content> = [];
 
-    // text content:
+    // text content. When the model refuses to answer, the API returns the
+    // refusal in `refusal` and sets `content` to null, so surface it as text
+    // instead of dropping it. Content and refusal are both returned when the
+    // model refuses after answering, so keep both:
     const text =
       choice.message.content != null && choice.message.content.length > 0
         ? choice.message.content
@@ -390,6 +402,12 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
       content.push({ type: 'text', text });
     }
 
+    const refusal = choice.message.refusal;
+    const hasRefusal = refusal != null && refusal.length > 0;
+    if (hasRefusal) {
+      content.push({ type: 'text', text: refusal });
+    }
+
     // tool calls:
     for (const toolCall of choice.message.tool_calls ?? []) {
       content.push({
@@ -423,14 +441,25 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
       providerMetadata.openai.rejectedPredictionTokens =
         completionTokenDetails?.rejected_prediction_tokens;
     }
-    if (choice.logprobs?.content != null) {
-      providerMetadata.openai.logprobs = choice.logprobs.content;
+    // logprobs either follow the text content (content) or the refusal, so
+    // combine them in the order the text is emitted above. Explicitly empty
+    // logprobs are reported as an empty array, as before:
+    const contentLogprobs = choice.logprobs?.content;
+    const refusalLogprobs = choice.logprobs?.refusal;
+    if (contentLogprobs != null || refusalLogprobs != null) {
+      providerMetadata.openai.logprobs = [
+        ...(contentLogprobs ?? []),
+        ...(refusalLogprobs ?? []),
+      ];
     }
 
     return {
       content,
       finishReason: {
-        unified: mapOpenAIFinishReason(choice.finish_reason),
+        unified: mapOpenAIFinishReason({
+          finishReason: choice.finish_reason,
+          hasRefusal,
+        }),
         raw: choice.finish_reason ?? undefined,
       },
       usage: convertOpenAIChatUsage(response.usage),
@@ -493,6 +522,13 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
     let usage: OpenAIChatUsage | undefined = undefined;
     let metadataExtracted = false;
     let isActiveText = false;
+    let hasRefusal = false;
+    let rawFinishReason: string | undefined = undefined;
+    let hasLogprobs = false;
+
+    // logprobs of all streamed chunks (content token logprobs followed by
+    // refusal token logprobs), collected across chunks:
+    const streamedLogprobs: OpenAIChatLogprob[] = [];
 
     const providerMetadata: SharedV3ProviderMetadata = { openai: {} };
 
@@ -518,6 +554,7 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
             // handle failed chunk parsing / validation:
             if (!chunk.success) {
               finishReason = { unified: 'error', raw: undefined };
+              rawFinishReason = undefined;
               controller.enqueue({ type: 'error', error: chunk.error });
               return;
             }
@@ -527,6 +564,7 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
             // handle error chunks:
             if ('error' in value) {
               finishReason = { unified: 'error', raw: undefined };
+              rawFinishReason = undefined;
               controller.enqueue({ type: 'error', error: value.error });
               return;
             }
@@ -566,15 +604,29 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
 
             const choice = value.choices[0];
 
+            // store the raw finish reason and map it in `flush`, after all
+            // chunks have been read: some servers send the last delta (which
+            // can contain a refusal) in the same chunk as the finish reason,
+            // and a refusal turns `stop` into `content-filter`:
             if (choice?.finish_reason != null) {
-              finishReason = {
-                unified: mapOpenAIFinishReason(choice.finish_reason),
-                raw: choice.finish_reason,
-              };
+              rawFinishReason = choice.finish_reason;
             }
 
-            if (choice?.logprobs?.content != null) {
-              providerMetadata.openai.logprobs = choice.logprobs.content;
+            // logprobs either follow the text content (content) or the refusal.
+            // They are streamed token by token, so collect the logprobs of all
+            // chunks (content logprobs followed by refusal logprobs) and report
+            // the complete list at the end of the stream:
+            if (choice?.logprobs != null) {
+              const contentLogprobs = choice.logprobs.content;
+              const refusalLogprobs = choice.logprobs.refusal;
+
+              if (contentLogprobs != null || refusalLogprobs != null) {
+                hasLogprobs = true;
+                streamedLogprobs.push(
+                  ...(contentLogprobs ?? []),
+                  ...(refusalLogprobs ?? []),
+                );
+              }
             }
 
             if (choice?.delta == null) {
@@ -583,7 +635,24 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
 
             const delta = choice.delta;
 
-            if (delta.content != null) {
+            // When the model refuses to answer, the API streams the refusal in
+            // `refusal` instead of `content`. Content and refusal deltas can
+            // both occur, so append each to the same text part instead of
+            // dropping one:
+            const textDeltas: string[] = [];
+            if (delta.content != null && delta.content.length > 0) {
+              textDeltas.push(delta.content);
+            }
+            if (delta.refusal != null && delta.refusal.length > 0) {
+              hasRefusal = true;
+              textDeltas.push(delta.refusal);
+            }
+            // empty `content: ""` deltas keep their previous behavior:
+            if (textDeltas.length === 0 && delta.content === '') {
+              textDeltas.push(delta.content);
+            }
+
+            for (const textDelta of textDeltas) {
               if (!isActiveText) {
                 controller.enqueue({ type: 'text-start', id: '0' });
                 isActiveText = true;
@@ -592,7 +661,7 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
               controller.enqueue({
                 type: 'text-delta',
                 id: '0',
-                delta: delta.content,
+                delta: textDelta,
               });
             }
 
@@ -623,6 +692,25 @@ export class OpenAIChatLanguageModel implements LanguageModelV3 {
 
             toolCallTracker.flush();
 
+            // map the stored finish reason now that all deltas (including
+            // refusals that arrived in the same chunk as the finish reason)
+            // have been read:
+            if (rawFinishReason != null) {
+              finishReason = {
+                unified: mapOpenAIFinishReason({
+                  finishReason: rawFinishReason,
+                  hasRefusal,
+                }),
+                raw: rawFinishReason,
+              };
+            }
+
+            // report the logprobs of all streamed chunks (content logprobs
+            // followed by refusal logprobs), not only the last chunk:
+            if (hasLogprobs) {
+              providerMetadata.openai.logprobs = streamedLogprobs;
+            }
+
             controller.enqueue({
               type: 'finish',
               finishReason,
@@ -650,6 +738,7 @@ function isOpenAIChatOutputChunk(chunk: OpenAIChatChunk): boolean {
 
     return (
       (delta?.content != null && delta.content.length > 0) ||
+      (delta?.refusal != null && delta.refusal.length > 0) ||
       (delta?.tool_calls != null && delta.tool_calls.length > 0) ||
       (delta?.annotations != null && delta.annotations.length > 0)
     );

package/src/responses/map-openai-responses-finish-reason.ts

@@ -3,15 +3,21 @@ import type { LanguageModelV3FinishReason } from '@ai-sdk/provider';
 export function mapOpenAIResponseFinishReason({
   finishReason,
   hasFunctionCall,
+  hasRefusal = false,
 }: {
   finishReason: string | null | undefined;
   // flag that checks if there have been client-side tool calls (not executed by openai)
   hasFunctionCall: boolean;
+  // flag that checks if the model refused to answer:
+  hasRefusal?: boolean;
 }): LanguageModelV3FinishReason['unified'] {
   switch (finishReason) {
     case undefined:
     case null:
-      return hasFunctionCall ? 'tool-calls' : 'stop';
+      // A client-side function call must win over a refusal so that the tools
+      // are still executed. The refusal text remains part of the content.
+      if (hasFunctionCall) return 'tool-calls';
+      return hasRefusal ? 'content-filter' : 'stop';
     case 'max_output_tokens':
       return 'length';
     case 'content_filter':

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

@@ -681,6 +681,8 @@ const openaiResponsesModeledChunkTypes = new Set([
   'response.output_item.done',
   'response.output_text.annotation.added',
   'response.output_text.delta',
+  'response.refusal.delta',
+  'response.refusal.done',
   'response.reasoning_summary_part.added',
   'response.reasoning_summary_part.done',
   'response.reasoning_summary_text.delta',
@@ -756,6 +758,20 @@ export const openaiResponsesChunkSchema = lazySchema(() =>
           )
           .nullish(),
       }),
+      z.object({
+        type: z.literal('response.refusal.delta'),
+        item_id: z.string(),
+        output_index: z.number().nullish(),
+        content_index: z.number().nullish(),
+        delta: z.string(),
+      }),
+      z.object({
+        type: z.literal('response.refusal.done'),
+        item_id: z.string(),
+        output_index: z.number().nullish(),
+        content_index: z.number().nullish(),
+        refusal: z.string(),
+      }),
       z.object({
         type: z.enum(['response.completed', 'response.incomplete']),
         response: z.object({
@@ -1330,55 +1346,62 @@ export const openaiResponsesResponseSchema = lazySchema(() =>
               role: z.literal('assistant'),
               id: z.string(),
               phase: z.enum(['commentary', 'final_answer']).nullish(),
+              // the API returns either text or a refusal as content parts:
               content: z.array(
-                z.object({
-                  type: z.literal('output_text'),
-                  text: z.string(),
-                  logprobs: z
-                    .array(
-                      z.object({
-                        token: z.string(),
-                        logprob: z.number(),
-                        top_logprobs: z.array(
-                          z.object({
-                            token: z.string(),
-                            logprob: z.number(),
-                          }),
-                        ),
-                      }),
-                    )
-                    .nullish(),
-                  annotations: z.array(
-                    z.discriminatedUnion('type', [
-                      z.object({
-                        type: z.literal('url_citation'),
-                        start_index: z.number(),
-                        end_index: z.number(),
-                        url: z.string(),
-                        title: z.string(),
-                      }),
-                      z.object({
-                        type: z.literal('file_citation'),
-                        file_id: z.string(),
-                        filename: z.string(),
-                        index: z.number(),
-                      }),
-                      z.object({
-                        type: z.literal('container_file_citation'),
-                        container_id: z.string(),
-                        file_id: z.string(),
-                        filename: z.string(),
-                        start_index: z.number(),
-                        end_index: z.number(),
-                      }),
-                      z.object({
-                        type: z.literal('file_path'),
-                        file_id: z.string(),
-                        index: z.number(),
-                      }),
-                    ]),
-                  ),
-                }),
+                z.discriminatedUnion('type', [
+                  z.object({
+                    type: z.literal('output_text'),
+                    text: z.string(),
+                    logprobs: z
+                      .array(
+                        z.object({
+                          token: z.string(),
+                          logprob: z.number(),
+                          top_logprobs: z.array(
+                            z.object({
+                              token: z.string(),
+                              logprob: z.number(),
+                            }),
+                          ),
+                        }),
+                      )
+                      .nullish(),
+                    annotations: z.array(
+                      z.discriminatedUnion('type', [
+                        z.object({
+                          type: z.literal('url_citation'),
+                          start_index: z.number(),
+                          end_index: z.number(),
+                          url: z.string(),
+                          title: z.string(),
+                        }),
+                        z.object({
+                          type: z.literal('file_citation'),
+                          file_id: z.string(),
+                          filename: z.string(),
+                          index: z.number(),
+                        }),
+                        z.object({
+                          type: z.literal('container_file_citation'),
+                          container_id: z.string(),
+                          file_id: z.string(),
+                          filename: z.string(),
+                          start_index: z.number(),
+                          end_index: z.number(),
+                        }),
+                        z.object({
+                          type: z.literal('file_path'),
+                          file_id: z.string(),
+                          index: z.number(),
+                        }),
+                      ]),
+                    ),
+                  }),
+                  z.object({
+                    type: z.literal('refusal'),
+                    refusal: z.string(),
+                  }),
+                ]),
               ),
             }),
             z.object({

package/src/responses/openai-responses-language-model.ts

@@ -686,6 +686,9 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
     let hasFunctionCall = false;
     const hostedToolSearchCallIds: string[] = [];
 
+    // flag that checks if the model refused to answer:
+    let hasRefusal = false;
+
     // map response content to content array (defined when there is no error)
     for (const part of response.output!) {
       switch (part.type) {
@@ -842,6 +845,23 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
 
         case 'message': {
           for (const contentPart of part.content) {
+            if (contentPart.type === 'refusal') {
+              hasRefusal = true;
+
+              content.push({
+                type: 'text',
+                text: contentPart.refusal,
+                providerMetadata: {
+                  [providerOptionsName]: {
+                    itemId: part.id,
+                    ...(part.phase != null && { phase: part.phase }),
+                  } satisfies ResponsesTextProviderMetadata,
+                },
+              });
+
+              continue;
+            }
+
             if (
               options.providerOptions?.[providerOptionsName]?.logprobs &&
               contentPart.logprobs
@@ -1200,6 +1220,7 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
         unified: mapOpenAIResponseFinishReason({
           finishReason: response.incomplete_details?.reason,
           hasFunctionCall,
+          hasRefusal,
         }),
         raw: response.incomplete_details?.reason ?? undefined,
       },
@@ -1320,6 +1341,9 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
     // flag that checks if there have been client-side tool calls (not executed by openai)
     let hasFunctionCall = false;
 
+    // flag that checks if the model refused to answer:
+    let hasRefusal = false;
+
     const activeReasoning: Record<
       string,
       {
@@ -2265,6 +2289,16 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
               ) {
                 logprobs.push(value.logprobs);
               }
+            } else if (isRefusalDeltaChunk(value)) {
+              hasRefusal = true;
+
+              controller.enqueue({
+                type: 'text-delta',
+                id: value.item_id,
+                delta: value.delta,
+              });
+            } else if (isRefusalDoneChunk(value)) {
+              hasRefusal = true;
             } else if (value.type === 'response.reasoning_summary_part.added') {
               // the first reasoning start is pushed in isResponseOutputItemAddedReasoningChunk
               if (value.summary_index > 0) {
@@ -2350,6 +2384,7 @@ export class OpenAIResponsesLanguageModel implements LanguageModelV3 {
                   unified: mapOpenAIResponseFinishReason({
                     finishReason: value.response.incomplete_details?.reason,
                     hasFunctionCall,
+                    hasRefusal,
                   }),
                   raw: value.response.incomplete_details?.reason ?? undefined,
                 };
@@ -2525,6 +2560,18 @@ function isTextDeltaChunk(
   return chunk.type === 'response.output_text.delta';
 }
 
+function isRefusalDeltaChunk(
+  chunk: OpenAIResponsesChunk,
+): chunk is OpenAIResponsesChunk & { type: 'response.refusal.delta' } {
+  return chunk.type === 'response.refusal.delta';
+}
+
+function isRefusalDoneChunk(
+  chunk: OpenAIResponsesChunk,
+): chunk is OpenAIResponsesChunk & { type: 'response.refusal.done' } {
+  return chunk.type === 'response.refusal.done';
+}
+
 function isOpenAIChatCompletionChunk(value: unknown): boolean {
   const chunk = asRecord(value);