@ai-sdk/openai 3.0.136 → 3.0.138

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.
@@ -2755,6 +2823,8 @@ const { images } = await generateImage({
 
 | Model                    | Sizes                                  |
 | ------------------------ | -------------------------------------- |
+| `gpt-image-2.6-flare`    | Standard presets and custom dimensions |
+| `gpt-image-2.6-sunburst` | Standard presets and custom dimensions |
 | `gpt-image-2.5-flare`    | Standard presets and custom dimensions |
 | `gpt-image-2.5-sunburst` | Standard presets and custom dimensions |
 | `gpt-image-2`            | 1024x1024, 1536x1024, 1024x1536        |
@@ -2802,6 +2872,28 @@ const { image } = await generateImage({
 });
 ```
 
+The `gpt-image-2.6-flare` and `gpt-image-2.6-sunburst` models support these additional options:
+
+- **complexity** _'low' | 'medium' | 'high'_ - how much detail the model puts into the image (generation and editing). The API defaults to `high`.
+- **outputFormat: 'gif'** _(generation only)_ - generates an animated GIF. Use **frameCount** (2-64, API default 9) and **frameDurationMs** (20-2000, API default 200) to control the animation.
+- **markup** _string_ _(editing only)_ - a transparent PNG (HTTPS URL or base64 data URL) overlaid on the first input image to mark up the regions to change. Supported by `gpt-image-2.5-*` and `gpt-image-2.6-*` models. It cannot be combined with `mask`; providing both throws an error.
+
+```ts
+const { image } = await generateImage({
+  model: openai.image('gpt-image-2.6-flare'),
+  prompt: 'A red panda waving hello, flat illustration',
+  providerOptions: {
+    openai: {
+      outputFormat: 'gif',
+      frameCount: 9,
+      frameDurationMs: 200,
+    } satisfies OpenAIImageModelGenerationOptions,
+  },
+});
+```
+
+All image options also accept `safetyIdentifier` (max 128 characters) to help OpenAI detect abuse by your end users.
+
 For more on `generateImage()` see [Image Generation](/docs/ai-sdk-core/image-generation).
 
 OpenAI's image models return additional metadata in the response that can be

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/openai",
-  "version": "3.0.136",
+  "version": "3.0.138",
   "license": "Apache-2.0",
   "sideEffects": false,
   "main": "./dist/index.js",
@@ -44,8 +44,8 @@
     "tsup": "^8",
     "typescript": "5.8.3",
     "zod": "3.25.76",
-    "@vercel/ai-tsconfig": "0.0.0",
-    "@ai-sdk/test-server": "1.0.6"
+    "@ai-sdk/test-server": "1.0.6",
+    "@vercel/ai-tsconfig": "0.0.0"
   },
   "peerDependencies": {
     "zod": "^3.25.76 || ^4.1.8"

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/image/openai-image-model-options.ts

@@ -12,6 +12,8 @@ export type OpenAIImageModelId =
   | 'gpt-image-1-mini'
   | 'gpt-image-1.5'
   | 'gpt-image-2'
+  | 'gpt-image-2.6-flare'
+  | 'gpt-image-2.6-sunburst'
   | 'gpt-image-2.5-flare'
   | 'gpt-image-2.5-flare-2026-09-08'
   | 'gpt-image-2.5-sunburst'
@@ -27,6 +29,8 @@ export const modelMaxImagesPerCall: Record<OpenAIImageModelId, number> = {
   'gpt-image-1-mini': 10,
   'gpt-image-1.5': 10,
   'gpt-image-2': 10,
+  'gpt-image-2.6-flare': 10,
+  'gpt-image-2.6-sunburst': 10,
   'gpt-image-2.5-flare': 10,
   'gpt-image-2.5-flare-2026-09-08': 10,
   'gpt-image-2.5-sunburst': 10,
@@ -74,6 +78,12 @@ const baseImageModelOptionsObject = () =>
      */
     outputFormat: z.enum(['png', 'jpeg', 'webp']).optional(),
 
+    /**
+     * How much detail the model puts into the image. Supported by GPT Image
+     * 2.6 models. The API defaults to `high`.
+     */
+    complexity: z.enum(['low', 'medium', 'high']).optional(),
+
     /**
      * Compression level (0-100) for the generated image(s). Applies to the
      * `jpeg` and `webp` output formats.
@@ -85,6 +95,13 @@ const baseImageModelOptionsObject = () =>
      * to monitor and detect abuse.
      */
     user: z.string().optional(),
+
+    /**
+     * A stable identifier used to help detect users of your application that
+     * may be violating OpenAI's usage policies. Should be a hashed username
+     * or email address.
+     */
+    safetyIdentifier: z.string().max(128).optional(),
   });
 
 export const openaiImageModelOptions = lazySchema(() =>
@@ -105,6 +122,25 @@ export const openaiImageModelGenerationOptions = lazySchema(() =>
        */
       style: z.enum(['vivid', 'natural']).optional(),
 
+      /**
+       * Format in which the generated image(s) are returned. `gif` produces
+       * an animated GIF and is supported by GPT Image 2.6 models.
+       */
+      outputFormat: z.enum(['png', 'jpeg', 'webp', 'gif']).optional(),
+
+      /**
+       * Number of frames (2-64) of an animated GIF. Only applies to the `gif`
+       * output format. The API defaults to 9.
+       */
+      frameCount: z.number().int().min(2).max(64).optional(),
+
+      /**
+       * Duration of each frame in milliseconds (20-2000) of an animated GIF.
+       * The API rounds it to the nearest 10 ms. Only applies to the `gif`
+       * output format. The API defaults to 200.
+       */
+      frameDurationMs: z.number().int().min(20).max(2000).optional(),
+
       /**
        * Content moderation level for the generated image(s). `low` applies
        * less restrictive filtering.
@@ -125,6 +161,14 @@ export const openaiImageModelEditOptions = lazySchema(() =>
        * Fidelity of the output image(s) to the input image(s).
        */
       inputFidelity: z.enum(['high', 'low']).optional(),
+
+      /**
+       * A transparent PNG overlaid on the first input image to mark up the
+       * regions to change, as an HTTPS URL or a base64 data URL. It must match
+       * the dimensions of the first image. Supported by GPT Image 2.5 and 2.6
+       * models. Cannot be combined with a `mask`.
+       */
+      markup: z.string().optional(),
     }),
   ),
 );

package/src/image/openai-image-model.ts

@@ -1,7 +1,8 @@
-import type {
-  ImageModelV3,
-  ImageModelV3File,
-  SharedV3Warning,
+import {
+  InvalidArgumentError,
+  type ImageModelV3,
+  type ImageModelV3File,
+  type SharedV3Warning,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
@@ -86,6 +87,13 @@ export class OpenAIImageModel implements ImageModelV3 {
           schema: openaiImageModelEditOptions,
         })) ?? {};
 
+      if (mask != null && openaiOptions.markup != null) {
+        throw new InvalidArgumentError({
+          argument: 'providerOptions.openai.markup',
+          message: '`markup` cannot be combined with `mask`.',
+        });
+      }
+
       const { value: response, responseHeaders } = await postFormDataToApi({
         url: this.config.url({
           path: '/images/edits',
@@ -114,14 +122,20 @@ export class OpenAIImageModel implements ImageModelV3 {
             ),
           ),
           mask: mask != null ? await fileToBlob(mask) : undefined,
+          markup:
+            openaiOptions.markup != null
+              ? await markupToBlob(openaiOptions.markup)
+              : undefined,
           n,
           size,
           quality: openaiOptions.quality,
           background: openaiOptions.background,
           output_format: openaiOptions.outputFormat,
           output_compression: openaiOptions.outputCompression,
+          complexity: openaiOptions.complexity,
           input_fidelity: openaiOptions.inputFidelity,
           user: openaiOptions.user,
+          safety_identifier: openaiOptions.safetyIdentifier,
         }),
         failedResponseHandler: openaiFailedResponseHandler,
         successfulResponseHandler: createJsonResponseHandler(
@@ -193,7 +207,11 @@ export class OpenAIImageModel implements ImageModelV3 {
         moderation: openaiOptions.moderation,
         output_format: openaiOptions.outputFormat,
         output_compression: openaiOptions.outputCompression,
+        complexity: openaiOptions.complexity,
+        frame_count: openaiOptions.frameCount,
+        frame_duration_ms: openaiOptions.frameDurationMs,
         user: openaiOptions.user,
+        safety_identifier: openaiOptions.safetyIdentifier,
         ...(!hasDefaultResponseFormat(this.modelId)
           ? { response_format: 'b64_json' }
           : {}),
@@ -284,16 +302,38 @@ type OpenAIImageEditInput = {
   prompt?: string;
   image: Blob | Blob[];
   mask?: Blob;
+  markup?: Blob;
   n?: number;
   size?: `${number}x${number}`;
   quality?: OpenAIImageModelEditOptions['quality'];
   background?: OpenAIImageModelEditOptions['background'];
   output_format?: OpenAIImageModelEditOptions['outputFormat'];
   output_compression?: OpenAIImageModelEditOptions['outputCompression'];
+  complexity?: OpenAIImageModelEditOptions['complexity'];
   input_fidelity?: OpenAIImageModelEditOptions['inputFidelity'];
   user?: OpenAIImageModelEditOptions['user'];
+  safety_identifier?: OpenAIImageModelEditOptions['safetyIdentifier'];
 };
 
+async function markupToBlob(markup: string): Promise<Blob> {
+  const dataUrl = /^data:([^;,]*)(;base64)?,(.*)$/s.exec(markup);
+
+  if (dataUrl == null) {
+    return downloadBlob(markup);
+  }
+
+  if (dataUrl[2] == null) {
+    throw new InvalidArgumentError({
+      argument: 'providerOptions.openai.markup',
+      message: '`markup` data URLs must be base64 encoded.',
+    });
+  }
+
+  return new Blob([convertBase64ToUint8Array(dataUrl[3]) as BlobPart], {
+    type: dataUrl[1] || 'image/png',
+  });
+}
+
 async function fileToBlob(
   file: ImageModelV3File | undefined,
 ): Promise<Blob | undefined> {

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':