@ai-sdk/gateway 4.0.64 → 4.0.68

package/docs/00-ai-gateway.mdx

@@ -207,6 +207,12 @@ const { text } = await generateText({
 
 AI Gateway language models can also be used in the `streamText` function and support structured data generation with [`Output`](/docs/reference/ai-sdk-core/output) (see [AI SDK Core](/docs/ai-sdk-core)).
 
+When an upstream provider reports a well-formed error after streaming has
+started, AI SDK Core exposes it as a
+[`StreamProviderError`](/docs/reference/ai-sdk-errors/ai-stream-provider-error).
+Use its `type`, `code`, `statusCode`, and `isRetryable` metadata to classify
+the failure without inspecting the Gateway's serialized error payload.
+
 ## Text Batches
 
 <Note type="warning">
@@ -1017,9 +1023,9 @@ The Tako Search tool supports these optional configuration options:
   - **includeContents** _boolean_ - Inline each page's extracted full text in `content`. **articleContentMaxChars** caps it, defaulting to 30,000 characters (maximum 1,000,000).
 - **sources.data** _object_ - Configure data results:
   - **count** _number_ - Maximum data results, 1-20. Defaults to 5.
-  - **includeContents** _boolean_ - Inline each card's underlying rows in `content.dataset` as typed, unit-labeled columns.
+  - **includeContents** _boolean_ - Inline each card's underlying rows in `content.dataset` as typed, unit-labeled columns. This adds a data surcharge based on the row count and dataset source. To estimate cost, first search with `includeContents: false` and inspect `cards.content.export_pricing`. Because this applies to every returned card, use `count` and `maxRows` to control cost.
   - **contentFormat** _'json_compact' | 'json_records' | 'csv' | 'card_json'_ - Serialization for inlined card data. Defaults to `'json_compact'`.
-  - **maxRows** _number_ - Row cap for inlined card data. Omit to use your account's default inline cap (20 rows on the standard plan); maximum 2,000.
+  - **maxRows** _number_ - Row cap per result. Omit to use the allowance in `cards.content.export_pricing`. A data surcharge applies per 1,000 exported rows; lower values reduce cost.
   - **nodeIds** _string[]_ - Data Graph node IDs to prioritize. Up to 20.
   - **strict** _boolean_ - Only return cards matching `nodeIds`. Requires at least one `nodeIds` value.
 - **location** _object_ - End-user `{ latitude, longitude }` coordinates for localized results.

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@ai-sdk/gateway",
   "private": false,
-  "version": "4.0.64",
+  "version": "4.0.68",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -31,7 +31,7 @@
   },
   "dependencies": {
     "@ai-sdk/provider": "4.0.8",
-    "@ai-sdk/provider-utils": "5.0.30",
+    "@ai-sdk/provider-utils": "5.0.33",
     "@vercel/oidc": "3.2.0"
   },
   "devDependencies": {

package/src/errors/as-gateway-error.ts

@@ -59,6 +59,11 @@ export async function asGatewayError(
       defaultMessage: 'Gateway request failed',
       cause: error,
       authMethod,
+      isRetryable:
+        error.isRetryable &&
+        (error.statusCode == null || error.statusCode < 400)
+          ? true
+          : undefined,
     });
   }
 

package/src/errors/create-gateway-error.ts

@@ -27,12 +27,14 @@ export async function createGatewayErrorFromResponse({
   defaultMessage = 'Gateway request failed',
   cause,
   authMethod,
+  isRetryable,
 }: {
   response: unknown;
   statusCode: number;
   defaultMessage?: string;
   cause?: unknown;
   authMethod?: 'api-key' | 'oidc';
+  isRetryable?: boolean;
 }): Promise<GatewayError> {
   const parseResult = await safeValidateTypes({
     value: response,
@@ -55,6 +57,7 @@ export async function createGatewayErrorFromResponse({
       validationError: parseResult.error,
       cause,
       generationId: rawGenerationId,
+      isRetryable,
     });
   }
 

package/src/errors/gateway-response-error.ts

@@ -23,6 +23,7 @@ export class GatewayResponseError extends GatewayError {
     validationError,
     cause,
     generationId,
+    isRetryable,
   }: {
     message?: string;
     statusCode?: number;
@@ -30,8 +31,9 @@ export class GatewayResponseError extends GatewayError {
     validationError?: TypeValidationError;
     cause?: unknown;
     generationId?: string;
+    isRetryable?: boolean;
   } = {}) {
-    super({ message, statusCode, cause, generationId });
+    super({ message, statusCode, cause, generationId, isRetryable });
     this.response = response;
     this.validationError = validationError;
   }

package/src/gateway-image-model-settings.ts

@@ -13,6 +13,7 @@ export type GatewayImageModelId =
   | 'bytedance/seedream-4.5'
   | 'bytedance/seedream-5.0-lite'
   | 'bytedance/seedream-5.0-pro'
+  | 'meta/muse-image-1.0'
   | 'openai/gpt-image-1'
   | 'openai/gpt-image-1-mini'
   | 'openai/gpt-image-1.5'

package/src/gateway-language-model-batch.ts

@@ -1,5 +1,4 @@
 import {
-  APICallError,
   type Experimental_BatchLanguageModelV4 as BatchLanguageModelV4,
   type Experimental_BatchV4ItemResult as BatchV4ItemResult,
   type Experimental_BatchV4OperationOptions as BatchV4OperationOptions,
@@ -15,9 +14,10 @@ import {
   combineHeaders,
   convertAsyncIteratorToReadableStream,
   createJsonErrorResponseHandler,
+  createJsonLinesResponseHandler,
   createJsonResponseHandler,
   getErrorMessage,
-  parseJSON,
+  normalizeBatchRequestCounts,
   postJsonToApi,
   resolve,
   WORKFLOW_SERIALIZE,
@@ -190,7 +190,7 @@ export class GatewayBatchLanguageModel
       : undefined;
 
     try {
-      const { value: stream } = await postJsonToApi({
+      const { value: lines } = await postJsonToApi({
         url: this.getBatchUrl('results'),
         headers: combineHeaders(
           resolvedHeaders,
@@ -199,28 +199,9 @@ export class GatewayBatchLanguageModel
           await resolve(this.config.o11yHeaders),
         ),
         body: { batchId },
-        successfulResponseHandler: async ({
-          response,
-          url,
-          requestBodyValues,
-        }: {
-          url: string;
-          requestBodyValues: unknown;
-          response: Response;
-        }) => {
-          if (response.body == null) {
-            throw new APICallError({
-              message: 'Batch results response body is empty',
-              url,
-              requestBodyValues,
-              statusCode: response.status,
-            });
-          }
-          return {
-            value: response.body,
-            responseHeaders: Object.fromEntries([...response.headers]),
-          };
-        },
+        successfulResponseHandler: createJsonLinesResponseHandler(
+          gatewayBatchItemResultLineSchema,
+        ),
         failedResponseHandler: createJsonErrorResponseHandler({
           errorSchema: z.any(),
           errorToMessage: data => getErrorMessage(data) ?? 'unknown error',
@@ -230,7 +211,7 @@ export class GatewayBatchLanguageModel
       });
 
       return convertAsyncIteratorToReadableStream(
-        parseGatewayBatchResultLines(stream),
+        convertGatewayBatchResultLines(lines),
       );
     } catch (error) {
       if (isAbortOrTimeoutError(error)) {
@@ -342,7 +323,12 @@ function convertGatewayBatchStatus(body: {
   expiresAt?: string | null;
   providerMetadata?: Record<string, Record<string, unknown>> | null;
 }): BatchV4Status {
-  const requestCounts = convertGatewayBatchRequestCounts(body.requestCounts);
+  const requestCounts = normalizeBatchRequestCounts({
+    total: body.requestCounts?.total,
+    pending: body.requestCounts?.pending,
+    completed: body.requestCounts?.completed,
+    failed: body.requestCounts?.failed,
+  });
 
   return {
     status: body.status,
@@ -367,112 +353,27 @@ function convertGatewayBatchStatus(body: {
 }
 
 /**
- * The spec's `requestCounts` requires all four counters; the Gateway's
- * persisted descriptor allows partial counts. Only forward counts when the
- * full set is present rather than fabricating zeros.
- */
-function convertGatewayBatchRequestCounts(
-  counts:
-    | {
-        total?: number | null;
-        pending?: number | null;
-        completed?: number | null;
-        failed?: number | null;
-      }
-    | null
-    | undefined,
-): BatchV4Status['requestCounts'] | undefined {
-  if (
-    counts == null ||
-    typeof counts.total !== 'number' ||
-    typeof counts.pending !== 'number' ||
-    typeof counts.completed !== 'number' ||
-    typeof counts.failed !== 'number'
-  ) {
-    return undefined;
-  }
-
-  return {
-    total: counts.total,
-    pending: counts.pending,
-    completed: counts.completed,
-    failed: counts.failed,
-  };
-}
-
-/**
- * Incremental NDJSON line splitter for the batch results stream: buffers
- * partial lines across chunks and flushes a trailing line without a final
- * newline. Each non-empty line is one `BatchV4ItemResult` JSON object.
+ * Converts the minimally validated Gateway batch result lines. The Gateway
+ * already sanitizes the complete result objects server-side.
  *
- * @param stream - The raw NDJSON byte stream from the batch results route.
- * @yields One minimally-validated `BatchV4ItemResult` per non-empty line.
+ * @yields Each Gateway batch item result.
  */
-async function* parseGatewayBatchResultLines(
-  stream: ReadableStream<Uint8Array>,
+async function* convertGatewayBatchResultLines(
+  lines: AsyncIterable<unknown>,
 ): AsyncGenerator<BatchV4ItemResult<LanguageModelV4GenerateResult>> {
-  const reader = stream.getReader();
-  const decoder = new TextDecoder();
-  let buffer = '';
-  let finished = false;
-
-  try {
-    while (true) {
-      const { done, value } = await reader.read();
-
-      if (done) {
-        finished = true;
-        buffer += decoder.decode();
-        break;
-      }
-
-      buffer += decoder.decode(value, { stream: true });
-
-      let lineEnd = buffer.indexOf('\n');
-      while (lineEnd !== -1) {
-        const line = buffer.slice(0, lineEnd).replace(/\r$/, '');
-        buffer = buffer.slice(lineEnd + 1);
-
-        if (line.trim().length > 0) {
-          yield await parseGatewayBatchResultLine(line);
-        }
-
-        lineEnd = buffer.indexOf('\n');
+  for await (const line of lines) {
+    const item = line as BatchV4ItemResult<LanguageModelV4GenerateResult>;
+
+    // JSON carries `response.timestamp` as an ISO string; core expects a Date.
+    if (item.status === 'succeeded') {
+      const response = item.result?.response;
+      if (response !== undefined && typeof response.timestamp === 'string') {
+        response.timestamp = new Date(response.timestamp);
       }
     }
 
-    const finalLine = buffer.replace(/\r$/, '');
-    if (finalLine.trim().length > 0) {
-      yield await parseGatewayBatchResultLine(finalLine);
-    }
-  } finally {
-    if (!finished) {
-      await reader.cancel().catch(() => {});
-    }
-    reader.releaseLock();
-  }
-}
-
-async function parseGatewayBatchResultLine(
-  line: string,
-): Promise<BatchV4ItemResult<LanguageModelV4GenerateResult>> {
-  // Minimal validation (id + status); items pass through otherwise — the
-  // Gateway already sanitizes them server-side.
-  const parsed = await parseJSON({
-    text: line,
-    schema: gatewayBatchItemResultLineSchema,
-  });
-  const item =
-    parsed as unknown as BatchV4ItemResult<LanguageModelV4GenerateResult>;
-  // JSON carries `response.timestamp` as an ISO string; core expects a Date
-  // (`GeneratedFile`-style consumers call `.toISOString()`).
-  if (item.status === 'succeeded') {
-    const response = item.result?.response;
-    if (response !== undefined && typeof response.timestamp === 'string') {
-      response.timestamp = new Date(response.timestamp);
-    }
+    yield item;
   }
-  return item;
 }
 
 const gatewayBatchItemResultLineSchema = z

package/src/gateway-language-model-settings.ts

@@ -26,6 +26,7 @@ export type GatewayModelId =
   | 'alibaba/qwen3.7-plus'
   | 'alibaba/qwen3.8-2.4t-a95b'
   | 'alibaba/qwen3.8-27b'
+  | 'alibaba/qwen3.8-flash'
   | 'alibaba/qwen3.8-max'
   | 'amazon/nova-2-lite'
   | 'amazon/nova-lite'
@@ -47,7 +48,6 @@ export type GatewayModelId =
   | 'anthropic/claude-sonnet-4.6'
   | 'anthropic/claude-sonnet-5'
   | 'arcee-ai/trinity-large-thinking'
-  | 'arcee-ai/trinity-mini'
   | 'bytedance/seed-1.6'
   | 'bytedance/seed-1.8'
   | 'cohere/command-a'
@@ -83,6 +83,8 @@ export type GatewayModelId =
   | 'inception/mercury-2'
   | 'inception/mercury-coder-small'
   | 'inclusionai/ling-3.0-flash'
+  | 'inclusionai/ling-3.0-flash-fin'
+  | 'inclusionai/ling-3.0-flash-fin-free'
   | 'interfaze/interfaze-beta'
   | 'kwaipilot/kat-coder-air-v2.5'
   | 'kwaipilot/kat-coder-pro-v1'
@@ -103,13 +105,13 @@ export type GatewayModelId =
   | 'minimax/minimax-m2.5'
   | 'minimax/minimax-m2.5-highspeed'
   | 'minimax/minimax-m2.7'
+  | 'minimax/minimax-m2.7-free'
   | 'minimax/minimax-m2.7-highspeed'
   | 'minimax/minimax-m3'
+  | 'minimax/minimax-m3-free'
   | 'mistral/codestral'
   | 'mistral/devstral-2'
   | 'mistral/devstral-small-2'
-  | 'mistral/magistral-medium'
-  | 'mistral/magistral-small'
   | 'mistral/ministral-14b'
   | 'mistral/ministral-3b'
   | 'mistral/ministral-8b'
@@ -133,7 +135,6 @@ export type GatewayModelId =
   | 'nvidia/nemotron-3-super-120b-a12b'
   | 'nvidia/nemotron-3-ultra-550b-a55b'
   | 'nvidia/nemotron-3.5-lightning'
-  | 'nvidia/nemotron-3.5-lightning-free'
   | 'nvidia/nemotron-nano-12b-v2-vl'
   | 'nvidia/nemotron-nano-9b-v2'
   | 'openai/gpt-3.5-turbo'
@@ -148,7 +149,6 @@ export type GatewayModelId =
   | 'openai/gpt-4o-fast'
   | 'openai/gpt-4o-mini'
   | 'openai/gpt-4o-mini-fast'
-  | 'openai/gpt-4o-mini-search-preview'
   | 'openai/gpt-5'
   | 'openai/gpt-5-codex'
   | 'openai/gpt-5-fast'
@@ -184,10 +184,10 @@ export type GatewayModelId =
   | 'openai/gpt-5.6-terra-fast'
   | 'openai/gpt-oss-120b'
   | 'openai/gpt-oss-20b'
+  | 'openai/gpt-oss-safeguard-120b'
   | 'openai/gpt-oss-safeguard-20b'
   | 'openai/o1'
   | 'openai/o3'
-  | 'openai/o3-deep-research'
   | 'openai/o3-fast'
   | 'openai/o3-mini'
   | 'openai/o3-pro'
@@ -235,5 +235,6 @@ export type GatewayModelId =
   | 'zai/glm-5.2'
   | 'zai/glm-5.2-fast'
   | 'zai/glm-5.3'
+  | 'zai/glm-5.3-flash'
   | 'zai/glm-5v-turbo'
   | (string & {});

package/src/gateway-transcription-model-settings.ts

@@ -1,6 +1,8 @@
 export type GatewayTranscriptionModelId =
   | 'fish-audio/transcribe-1'
   | 'fish-audio/transcribe-1-free'
+  | 'google/gemini-3.5-transcribe'
+  | 'google/gemini-3.5-transcribe-live'
   | 'openai/gpt-4o-mini-transcribe'
   | 'openai/gpt-4o-transcribe'
   | 'openai/gpt-realtime-whisper'

package/src/gateway-video-model-settings.ts

@@ -7,6 +7,7 @@ export type GatewayVideoModelId =
   | 'alibaba/wan-v2.6-t2v'
   | 'alibaba/wan-v2.7-r2v'
   | 'alibaba/wan-v2.7-t2v'
+  | 'alibaba/wan-v3.0-video'
   | 'bfl/flux-3-video'
   | 'bytedance/seedance-2.0'
   | 'bytedance/seedance-2.0-fast'

package/src/tool/tako-search.ts

@@ -14,15 +14,28 @@ export type TakoContentFormat =
   | 'json_records';
 
 export interface TakoDataSourceConfig {
-  /** Maximum number of data cards to return (1-20). */
+  /**
+   * Maximum number of data results to return (1-20). When includeContents is
+   * true, each additional result adds its own data surcharge.
+   */
   count?: number;
-  /** Inline underlying card data in the response. This can add a data charge. */
+  /**
+   * Inline rows for each data result. This adds a data surcharge based on row
+   * count and dataset source. To estimate cost, search with includeContents
+   * disabled and inspect cards.content.export_pricing. This applies to every
+   * returned card; limit sources.data.count and sources.data.maxRows to control
+   * cost.
+   */
   includeContents?: boolean;
   /** Requested delivery mode for card data. Search cards are always inlined. */
   mode?: 'inline' | 'url';
   /** Serialization for inlined card data. */
   contentFormat?: TakoContentFormat;
-  /** Maximum rows for inlined card data. The service clamps large values. */
+  /**
+   * Maximum rows to inline per result. Omit to use the allowance in
+   * cards.content.export_pricing. A data surcharge applies per 1,000 exported
+   * rows; lower values reduce cost.
+   */
   maxRows?: number;
   /** Data Graph node IDs to prioritize. */
   nodeIds?: string[];
@@ -190,7 +203,6 @@ export interface TakoCard {
   relevance?: 'High' | 'Low' | 'Medium' | null;
   content?: TakoResultContent | null;
   exportable?: boolean;
-  relevance_score?: number | null;
   nodes?: Array<{
     id: string;
     type: 'entity' | 'metric';
@@ -215,7 +227,6 @@ export interface TakoWebResult {
   source_name?: string | null;
   publish_date?: string | null;
   content?: TakoResultContent | null;
-  citation_number?: number | null;
 }
 
 export interface TakoSearchResponse {
@@ -254,11 +265,15 @@ const takoDataSourceInputSchema = z.object({
   count: z
     .number()
     .optional()
-    .describe('Maximum number of data cards to return (1-20).'),
+    .describe(
+      'Maximum number of data results to return (1-20). When include_contents is true, each additional result adds its own data surcharge.',
+    ),
   include_contents: z
     .boolean()
     .optional()
-    .describe('Inline underlying card data. This can add a data charge.'),
+    .describe(
+      'Inline rows for each data result. This adds a data surcharge based on row count and dataset source. To estimate cost, search with include_contents disabled and inspect cards.content.export_pricing. This applies to every returned card; limit sources.data.count and sources.data.max_rows to control cost.',
+    ),
   mode: z
     .enum(['inline', 'url'])
     .optional()
@@ -272,7 +287,9 @@ const takoDataSourceInputSchema = z.object({
   max_rows: z
     .number()
     .optional()
-    .describe('Maximum rows for inlined card data.'),
+    .describe(
+      'Maximum rows to inline per result. Omit to use the allowance in cards.content.export_pricing. A data surcharge applies per 1,000 exported rows; lower values reduce cost.',
+    ),
   node_ids: z
     .array(z.string())
     .optional()
@@ -493,7 +510,6 @@ const takoCardSchema = z
     relevance: z.enum(['High', 'Low', 'Medium']).nullish(),
     content: takoResultContentSchema.nullish(),
     exportable: z.boolean().optional(),
-    relevance_score: z.number().nullish(),
     nodes: z
       .array(
         z.object({
@@ -525,7 +541,6 @@ const takoWebResultSchema = z
     source_name: z.string().nullish(),
     publish_date: z.string().nullish(),
     content: takoResultContentSchema.nullish(),
-    citation_number: z.number().nullish(),
   })
   .passthrough();