@ai-sdk/xai 4.0.0-beta.2 → 4.0.0-beta.21

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

@@ -1,19 +1,21 @@
 import {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3StreamPart,
-  LanguageModelV3StreamResult,
-  LanguageModelV3Usage,
-  SharedV3Warning,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4Content,
+  LanguageModelV4FinishReason,
+  LanguageModelV4GenerateResult,
+  LanguageModelV4StreamPart,
+  LanguageModelV4StreamResult,
+  LanguageModelV4Usage,
+  SharedV4Warning,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
   createEventSourceResponseHandler,
   createJsonResponseHandler,
   FetchFunction,
+  isCustomReasoning,
+  mapReasoningToProviderEffort,
   parseProviderOptions,
   ParseResult,
   postJsonToApi,
@@ -43,8 +45,8 @@ type XaiResponsesConfig = {
   fetch?: FetchFunction;
 };
 
-export class XaiResponsesLanguageModel implements LanguageModelV3 {
-  readonly specificationVersion = 'v3';
+export class XaiResponsesLanguageModel implements LanguageModelV4 {
+  readonly specificationVersion = 'v4';
 
   readonly modelId: XaiResponsesModelId;
 
@@ -74,8 +76,9 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
     providerOptions,
     tools,
     toolChoice,
-  }: LanguageModelV3CallOptions) {
-    const warnings: SharedV3Warning[] = [];
+    reasoning,
+  }: LanguageModelV4CallOptions) {
+    const warnings: SharedV4Warning[] = [];
 
     const options =
       (await parseProviderOptions({
@@ -139,6 +142,24 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
       }
     }
 
+    const resolvedReasoningEffort =
+      options.reasoningEffort ??
+      (isCustomReasoning(reasoning)
+        ? reasoning === 'none'
+          ? undefined
+          : mapReasoningToProviderEffort({
+              reasoning,
+              effortMap: {
+                minimal: 'low',
+                low: 'low',
+                medium: 'medium',
+                high: 'high',
+                xhigh: 'high',
+              },
+              warnings,
+            })
+        : undefined);
+
     const baseArgs: Record<string, unknown> = {
       model: this.modelId,
       input,
@@ -165,8 +186,16 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
               : { type: 'json_object' },
         },
       }),
-      ...(options.reasoningEffort != null && {
-        reasoning: { effort: options.reasoningEffort },
+      ...((resolvedReasoningEffort != null ||
+        options.reasoningSummary != null) && {
+        reasoning: {
+          ...(resolvedReasoningEffort != null && {
+            effort: resolvedReasoningEffort,
+          }),
+          ...(options.reasoningSummary != null && {
+            summary: options.reasoningSummary,
+          }),
+        },
       }),
       ...(options.store === false && {
         store: options.store,
@@ -199,8 +228,8 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
   }
 
   async doGenerate(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3GenerateResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4GenerateResult> {
     const {
       args: body,
       warnings,
@@ -227,7 +256,8 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
       fetch: this.config.fetch,
     });
 
-    const content: Array<LanguageModelV3Content> = [];
+    const content: Array<LanguageModelV4Content> = [];
+    let hasFunctionCall = false;
 
     const webSearchSubTools = [
       'web_search',
@@ -350,6 +380,7 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
         }
 
         case 'function_call': {
+          hasFunctionCall = true;
           content.push({
             type: 'tool-call',
             toolCallId: part.call_id,
@@ -360,12 +391,16 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
         }
 
         case 'reasoning': {
-          const summaryTexts = part.summary
-            .map(s => s.text)
-            .filter(text => text && text.length > 0);
+          const texts =
+            part.summary.length > 0
+              ? part.summary.map(s => s.text)
+              : (part.content ?? []).map(c => c.text);
+
+          const reasoningText = texts
+            .filter(text => text && text.length > 0)
+            .join('');
 
-          if (summaryTexts.length > 0) {
-            const reasoningText = summaryTexts.join('');
+          if (reasoningText) {
             if (part.encrypted_content || part.id) {
               content.push({
                 type: 'reasoning',
@@ -398,7 +433,9 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
     return {
       content,
       finishReason: {
-        unified: mapXaiResponsesFinishReason(response.status),
+        unified: hasFunctionCall
+          ? 'tool-calls'
+          : mapXaiResponsesFinishReason(response.status),
         raw: response.status ?? undefined,
       },
       usage: response.usage
@@ -418,8 +455,8 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
   }
 
   async doStream(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3StreamResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4StreamResult> {
     const {
       args,
       warnings,
@@ -446,11 +483,12 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
       fetch: this.config.fetch,
     });
 
-    let finishReason: LanguageModelV3FinishReason = {
+    let finishReason: LanguageModelV4FinishReason = {
       unified: 'other',
       raw: undefined,
     };
-    let usage: LanguageModelV3Usage | undefined = undefined;
+    let hasFunctionCall = false;
+    let usage: LanguageModelV4Usage | undefined = undefined;
     let isFirstChunk = true;
     const contentBlocks: Record<string, { type: 'text' }> = {};
     const seenToolCalls = new Set<string>();
@@ -473,7 +511,7 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
       stream: response.pipeThrough(
         new TransformStream<
           ParseResult<z.infer<typeof xaiResponsesChunkSchema>>,
-          LanguageModelV3StreamPart
+          LanguageModelV4StreamPart
         >({
           start(controller) {
             controller.enqueue({ type: 'stream-start', warnings });
@@ -643,7 +681,9 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
 
               if (response.status) {
                 finishReason = {
-                  unified: mapXaiResponsesFinishReason(response.status),
+                  unified: hasFunctionCall
+                    ? 'tool-calls'
+                    : mapXaiResponsesFinishReason(response.status),
                   raw: response.status,
                 };
               }
@@ -911,6 +951,7 @@ export class XaiResponsesLanguageModel implements LanguageModelV3 {
                     toolName: part.name,
                   });
                 } else if (event.type === 'response.output_item.done') {
+                  hasFunctionCall = true;
                   ongoingToolCalls[event.output_index] = undefined;
 
                   controller.enqueue({

package/src/responses/xai-responses-options.ts

@@ -6,6 +6,9 @@ export type XaiResponsesModelId =
   | 'grok-4'
   | 'grok-4-fast-non-reasoning'
   | 'grok-4-fast-reasoning'
+  | 'grok-4.20-0309-non-reasoning'
+  | 'grok-4.20-0309-reasoning'
+  | 'grok-4.20-multi-agent-0309'
   | (string & {});
 
 /**
@@ -17,6 +20,7 @@ export const xaiLanguageModelResponsesOptions = z.object({
    * Possible values are `low` (uses fewer reasoning tokens), `medium` and `high` (uses more reasoning tokens).
    */
   reasoningEffort: z.enum(['low', 'medium', 'high']).optional(),
+  reasoningSummary: z.enum(['auto', 'concise', 'detailed']).optional(),
   logprobs: z.boolean().optional(),
   topLogprobs: z.number().int().min(0).max(8).optional(),
   /**

package/src/responses/xai-responses-prepare-tools.ts

@@ -1,6 +1,6 @@
 import {
-  LanguageModelV3CallOptions,
-  SharedV3Warning,
+  LanguageModelV4CallOptions,
+  SharedV4Warning,
   UnsupportedFunctionalityError,
 } from '@ai-sdk/provider';
 import { validateTypes } from '@ai-sdk/provider-utils';
@@ -20,16 +20,16 @@ export async function prepareResponsesTools({
   tools,
   toolChoice,
 }: {
-  tools: LanguageModelV3CallOptions['tools'];
-  toolChoice?: LanguageModelV3CallOptions['toolChoice'];
+  tools: LanguageModelV4CallOptions['tools'];
+  toolChoice?: LanguageModelV4CallOptions['toolChoice'];
 }): Promise<{
   tools: Array<XaiResponsesTool> | undefined;
   toolChoice: XaiResponsesToolChoice | undefined;
-  toolWarnings: SharedV3Warning[];
+  toolWarnings: SharedV4Warning[];
 }> {
   const normalizedTools = tools?.length ? tools : undefined;
 
-  const toolWarnings: SharedV3Warning[] = [];
+  const toolWarnings: SharedV4Warning[] = [];
 
   if (normalizedTools == null) {
     return { tools: undefined, toolChoice: undefined, toolWarnings };

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

@@ -1,14 +1,14 @@
 import {
   APICallError,
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  LanguageModelV3Content,
-  LanguageModelV3FinishReason,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3StreamPart,
-  LanguageModelV3StreamResult,
-  LanguageModelV3Usage,
-  SharedV3Warning,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4Content,
+  LanguageModelV4FinishReason,
+  LanguageModelV4GenerateResult,
+  LanguageModelV4StreamPart,
+  LanguageModelV4StreamResult,
+  LanguageModelV4Usage,
+  SharedV4Warning,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
@@ -16,6 +16,8 @@ import {
   createJsonResponseHandler,
   extractResponseHeaders,
   FetchFunction,
+  isCustomReasoning,
+  mapReasoningToProviderEffort,
   parseProviderOptions,
   ParseResult,
   postJsonToApi,
@@ -41,8 +43,8 @@ type XaiChatConfig = {
   fetch?: FetchFunction;
 };
 
-export class XaiChatLanguageModel implements LanguageModelV3 {
-  readonly specificationVersion = 'v3';
+export class XaiChatLanguageModel implements LanguageModelV4 {
+  readonly specificationVersion = 'v4';
 
   readonly modelId: XaiChatModelId;
 
@@ -71,12 +73,13 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
     presencePenalty,
     stopSequences,
     seed,
+    reasoning,
     responseFormat,
     providerOptions,
     tools,
     toolChoice,
-  }: LanguageModelV3CallOptions) {
-    const warnings: SharedV3Warning[] = [];
+  }: LanguageModelV4CallOptions) {
+    const warnings: SharedV4Warning[] = [];
 
     // parse xai-specific provider options
     const options =
@@ -133,7 +136,23 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
       temperature,
       top_p: topP,
       seed,
-      reasoning_effort: options.reasoningEffort,
+      reasoning_effort:
+        options.reasoningEffort ??
+        (isCustomReasoning(reasoning)
+          ? reasoning === 'none'
+            ? undefined
+            : mapReasoningToProviderEffort({
+                reasoning,
+                effortMap: {
+                  minimal: 'low',
+                  low: 'low',
+                  medium: 'low',
+                  high: 'high',
+                  xhigh: 'high',
+                },
+                warnings,
+              })
+          : undefined),
 
       // parallel function calling
       parallel_function_calling: options.parallel_function_calling,
@@ -202,8 +221,8 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
   }
 
   async doGenerate(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3GenerateResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4GenerateResult> {
     const { args: body, warnings } = await this.getArgs(options);
 
     const url = `${this.config.baseURL ?? 'https://api.x.ai/v1'}/chat/completions`;
@@ -237,7 +256,7 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
     }
 
     const choice = response.choices![0];
-    const content: Array<LanguageModelV3Content> = [];
+    const content: Array<LanguageModelV4Content> = [];
 
     // extract text content
     if (choice.message.content != null && choice.message.content.length > 0) {
@@ -312,8 +331,8 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
   }
 
   async doStream(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3StreamResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4StreamResult> {
     const { args, warnings } = await this.getArgs(options);
     const body = {
       ...args,
@@ -375,11 +394,11 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
       fetch: this.config.fetch,
     });
 
-    let finishReason: LanguageModelV3FinishReason = {
+    let finishReason: LanguageModelV4FinishReason = {
       unified: 'other',
       raw: undefined,
     };
-    let usage: LanguageModelV3Usage | undefined = undefined;
+    let usage: LanguageModelV4Usage | undefined = undefined;
     let isFirstChunk = true;
     const contentBlocks: Record<
       string,
@@ -394,7 +413,7 @@ export class XaiChatLanguageModel implements LanguageModelV3 {
       stream: response.pipeThrough(
         new TransformStream<
           ParseResult<z.infer<typeof xaiChatChunkSchema>>,
-          LanguageModelV3StreamPart
+          LanguageModelV4StreamPart
         >({
           start(controller) {
             controller.enqueue({ type: 'stream-start', warnings });

package/src/xai-chat-options.ts

@@ -6,6 +6,9 @@ export type XaiChatModelId =
   | 'grok-4-1-fast-non-reasoning'
   | 'grok-4-fast-non-reasoning'
   | 'grok-4-fast-reasoning'
+  | 'grok-4.20-0309-non-reasoning'
+  | 'grok-4.20-0309-reasoning'
+  | 'grok-4.20-multi-agent-0309'
   | 'grok-code-fast-1'
   | 'grok-4'
   | 'grok-4-0709'
@@ -14,12 +17,6 @@ export type XaiChatModelId =
   | 'grok-3-latest'
   | 'grok-3-mini'
   | 'grok-3-mini-latest'
-  | 'grok-2-vision-1212'
-  | 'grok-2-vision'
-  | 'grok-2-vision-latest'
-  | 'grok-2-image-1212'
-  | 'grok-2-image'
-  | 'grok-2-image-latest'
   | (string & {});
 
 // search source schemas

package/src/xai-image-model.ts

@@ -1,4 +1,4 @@
-import { ImageModelV3, SharedV3Warning } from '@ai-sdk/provider';
+import { ImageModelV4, SharedV4Warning } from '@ai-sdk/provider';
 import {
   combineHeaders,
   convertImageModelFileToDataUri,
@@ -25,9 +25,9 @@ interface XaiImageModelConfig {
   };
 }
 
-export class XaiImageModel implements ImageModelV3 {
-  readonly specificationVersion = 'v3';
-  readonly maxImagesPerCall = 1;
+export class XaiImageModel implements ImageModelV4 {
+  readonly specificationVersion = 'v4';
+  readonly maxImagesPerCall = 3;
 
   get provider(): string {
     return this.config.provider;
@@ -49,10 +49,10 @@ export class XaiImageModel implements ImageModelV3 {
     abortSignal,
     files,
     mask,
-  }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<
-    Awaited<ReturnType<ImageModelV3['doGenerate']>>
+  }: Parameters<ImageModelV4['doGenerate']>[0]): Promise<
+    Awaited<ReturnType<ImageModelV4['doGenerate']>>
   > {
-    const warnings: Array<SharedV3Warning> = [];
+    const warnings: Array<SharedV4Warning> = [];
 
     if (size != null) {
       warnings.push({
@@ -84,19 +84,9 @@ export class XaiImageModel implements ImageModelV3 {
     });
 
     const hasFiles = files != null && files.length > 0;
-    let imageUrl: string | undefined;
-
-    if (hasFiles) {
-      imageUrl = convertImageModelFileToDataUri(files[0]);
-
-      if (files.length > 1) {
-        warnings.push({
-          type: 'other',
-          message:
-            'xAI only supports a single input image. Additional images are ignored.',
-        });
-      }
-    }
+    const imageUrls = hasFiles
+      ? files.map(file => convertImageModelFileToDataUri(file))
+      : [];
 
     const endpoint = hasFiles ? '/images/edits' : '/images/generations';
 
@@ -104,7 +94,7 @@ export class XaiImageModel implements ImageModelV3 {
       model: this.modelId,
       prompt,
       n,
-      response_format: 'url',
+      response_format: 'b64_json',
     };
 
     if (aspectRatio != null) {
@@ -127,8 +117,18 @@ export class XaiImageModel implements ImageModelV3 {
       body.resolution = xaiOptions.resolution;
     }
 
-    if (imageUrl != null) {
-      body.image = { url: imageUrl, type: 'image_url' };
+    if (xaiOptions?.quality != null) {
+      body.quality = xaiOptions.quality;
+    }
+
+    if (xaiOptions?.user != null) {
+      body.user = xaiOptions.user;
+    }
+
+    if (imageUrls.length === 1) {
+      body.image = { url: imageUrls[0], type: 'image_url' };
+    } else if (imageUrls.length > 1) {
+      body.images = imageUrls.map(url => ({ url, type: 'image_url' }));
     }
 
     const baseURL = this.config.baseURL ?? 'https://api.x.ai/v1';
@@ -145,12 +145,18 @@ export class XaiImageModel implements ImageModelV3 {
       fetch: this.config.fetch,
     });
 
-    const downloadedImages = await Promise.all(
-      response.data.map(image => this.downloadImage(image.url, abortSignal)),
-    );
+    const hasAllBase64 = response.data.every(image => image.b64_json != null);
+
+    const images = hasAllBase64
+      ? response.data.map(image => image.b64_json!)
+      : await Promise.all(
+          response.data.map(image =>
+            this.downloadImage(image.url!, abortSignal),
+          ),
+        );
 
     return {
-      images: downloadedImages,
+      images,
       warnings,
       response: {
         timestamp: currentDate,
@@ -164,6 +170,9 @@ export class XaiImageModel implements ImageModelV3 {
               ? { revisedPrompt: item.revised_prompt }
               : {}),
           })),
+          ...(response.usage?.cost_in_usd_ticks != null
+            ? { costInUsdTicks: response.usage.cost_in_usd_ticks }
+            : {}),
         },
       },
     };
@@ -187,8 +196,14 @@ export class XaiImageModel implements ImageModelV3 {
 const xaiImageResponseSchema = z.object({
   data: z.array(
     z.object({
-      url: z.string(),
+      url: z.string().nullish(),
+      b64_json: z.string().nullish(),
       revised_prompt: z.string().nullish(),
     }),
   ),
+  usage: z
+    .object({
+      cost_in_usd_ticks: z.number().nullish(),
+    })
+    .nullish(),
 });

package/src/xai-image-options.ts

@@ -5,6 +5,8 @@ export const xaiImageModelOptions = z.object({
   output_format: z.string().optional(),
   sync_mode: z.boolean().optional(),
   resolution: z.enum(['1k', '2k']).optional(),
+  quality: z.enum(['low', 'medium', 'high']).optional(),
+  user: z.string().optional(),
 });
 
 export type XaiImageModelOptions = z.infer<typeof xaiImageModelOptions>;

package/src/xai-image-settings.ts

@@ -1,6 +1,4 @@
 export type XaiImageModelId =
-  | 'grok-2-image'
-  | 'grok-2-image-1212'
   | 'grok-imagine-image'
   | 'grok-imagine-image-pro'
   | (string & {});

package/src/xai-prepare-tools.ts

@@ -1,6 +1,6 @@
 import {
-  LanguageModelV3CallOptions,
-  SharedV3Warning,
+  LanguageModelV4CallOptions,
+  SharedV4Warning,
   UnsupportedFunctionalityError,
 } from '@ai-sdk/provider';
 import { XaiToolChoice } from './xai-chat-prompt';
@@ -9,8 +9,8 @@ export function prepareTools({
   tools,
   toolChoice,
 }: {
-  tools: LanguageModelV3CallOptions['tools'];
-  toolChoice?: LanguageModelV3CallOptions['toolChoice'];
+  tools: LanguageModelV4CallOptions['tools'];
+  toolChoice?: LanguageModelV4CallOptions['toolChoice'];
 }): {
   tools:
     | Array<{
@@ -24,12 +24,12 @@ export function prepareTools({
       }>
     | undefined;
   toolChoice: XaiToolChoice | undefined;
-  toolWarnings: SharedV3Warning[];
+  toolWarnings: SharedV4Warning[];
 } {
   // when the tools array is empty, change it to undefined to prevent errors
   tools = tools?.length ? tools : undefined;
 
-  const toolWarnings: SharedV3Warning[] = [];
+  const toolWarnings: SharedV4Warning[] = [];
 
   if (tools == null) {
     return { tools: undefined, toolChoice: undefined, toolWarnings };