@ai-sdk/openai 3.0.0-beta.102 → 3.0.0-beta.105

package/dist/internal/index.d.mts

@@ -19,6 +19,8 @@ declare const openaiChatLanguageModelOptions: _ai_sdk_provider_utils.LazySchema<
     promptCacheKey?: string | undefined;
     promptCacheRetention?: "in_memory" | "24h" | undefined;
     safetyIdentifier?: string | undefined;
+    systemMessageMode?: "remove" | "system" | "developer" | undefined;
+    forceReasoning?: boolean | undefined;
 }>;
 type OpenAIChatLanguageModelOptions = InferSchema<typeof openaiChatLanguageModelOptions>;
 
@@ -115,7 +117,7 @@ declare class OpenAIEmbeddingModel implements EmbeddingModelV3 {
     doEmbed({ values, headers, abortSignal, providerOptions, }: Parameters<EmbeddingModelV3['doEmbed']>[0]): Promise<Awaited<ReturnType<EmbeddingModelV3['doEmbed']>>>;
 }
 
-type OpenAIImageModelId = 'dall-e-3' | 'dall-e-2' | 'gpt-image-1' | 'gpt-image-1-mini' | (string & {});
+type OpenAIImageModelId = 'dall-e-3' | 'dall-e-2' | 'gpt-image-1' | 'gpt-image-1-mini' | 'gpt-image-1.5' | (string & {});
 declare const modelMaxImagesPerCall: Record<OpenAIImageModelId, number>;
 declare const hasDefaultResponseFormat: Set<string>;
 
@@ -131,7 +133,7 @@ declare class OpenAIImageModel implements ImageModelV3 {
     get maxImagesPerCall(): number;
     get provider(): string;
     constructor(modelId: OpenAIImageModelId, config: OpenAIImageModelConfig);
-    doGenerate({ prompt, n, size, aspectRatio, seed, providerOptions, headers, abortSignal, }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<Awaited<ReturnType<ImageModelV3['doGenerate']>>>;
+    doGenerate({ prompt, files, mask, n, size, aspectRatio, seed, providerOptions, headers, abortSignal, }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<Awaited<ReturnType<ImageModelV3['doGenerate']>>>;
 }
 
 type OpenAITranscriptionModelId = 'whisper-1' | 'gpt-4o-mini-transcribe' | 'gpt-4o-transcribe' | (string & {});
@@ -525,7 +527,7 @@ declare const fileSearch: _ai_sdk_provider_utils.ProviderToolFactoryWithOutputSc
 }>;
 
 declare const imageGenerationArgsSchema: _ai_sdk_provider_utils.LazySchema<{
-    background?: "auto" | "opaque" | "transparent" | undefined;
+    background?: "auto" | "transparent" | "opaque" | undefined;
     inputFidelity?: "low" | "high" | undefined;
     inputImageMask?: {
         fileId?: string | undefined;

package/dist/internal/index.d.ts

@@ -19,6 +19,8 @@ declare const openaiChatLanguageModelOptions: _ai_sdk_provider_utils.LazySchema<
     promptCacheKey?: string | undefined;
     promptCacheRetention?: "in_memory" | "24h" | undefined;
     safetyIdentifier?: string | undefined;
+    systemMessageMode?: "remove" | "system" | "developer" | undefined;
+    forceReasoning?: boolean | undefined;
 }>;
 type OpenAIChatLanguageModelOptions = InferSchema<typeof openaiChatLanguageModelOptions>;
 
@@ -115,7 +117,7 @@ declare class OpenAIEmbeddingModel implements EmbeddingModelV3 {
     doEmbed({ values, headers, abortSignal, providerOptions, }: Parameters<EmbeddingModelV3['doEmbed']>[0]): Promise<Awaited<ReturnType<EmbeddingModelV3['doEmbed']>>>;
 }
 
-type OpenAIImageModelId = 'dall-e-3' | 'dall-e-2' | 'gpt-image-1' | 'gpt-image-1-mini' | (string & {});
+type OpenAIImageModelId = 'dall-e-3' | 'dall-e-2' | 'gpt-image-1' | 'gpt-image-1-mini' | 'gpt-image-1.5' | (string & {});
 declare const modelMaxImagesPerCall: Record<OpenAIImageModelId, number>;
 declare const hasDefaultResponseFormat: Set<string>;
 
@@ -131,7 +133,7 @@ declare class OpenAIImageModel implements ImageModelV3 {
     get maxImagesPerCall(): number;
     get provider(): string;
     constructor(modelId: OpenAIImageModelId, config: OpenAIImageModelConfig);
-    doGenerate({ prompt, n, size, aspectRatio, seed, providerOptions, headers, abortSignal, }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<Awaited<ReturnType<ImageModelV3['doGenerate']>>>;
+    doGenerate({ prompt, files, mask, n, size, aspectRatio, seed, providerOptions, headers, abortSignal, }: Parameters<ImageModelV3['doGenerate']>[0]): Promise<Awaited<ReturnType<ImageModelV3['doGenerate']>>>;
 }
 
 type OpenAITranscriptionModelId = 'whisper-1' | 'gpt-4o-mini-transcribe' | 'gpt-4o-transcribe' | (string & {});
@@ -525,7 +527,7 @@ declare const fileSearch: _ai_sdk_provider_utils.ProviderToolFactoryWithOutputSc
 }>;
 
 declare const imageGenerationArgsSchema: _ai_sdk_provider_utils.LazySchema<{
-    background?: "auto" | "opaque" | "transparent" | undefined;
+    background?: "auto" | "transparent" | "opaque" | undefined;
     inputFidelity?: "low" | "high" | undefined;
     inputImageMask?: {
         fileId?: string | undefined;

package/dist/internal/index.js

@@ -83,7 +83,7 @@ var openaiFailedResponseHandler = (0, import_provider_utils.createJsonErrorRespo
 function getOpenAILanguageModelCapabilities(modelId) {
   const supportsFlexProcessing = modelId.startsWith("o3") || modelId.startsWith("o4-mini") || modelId.startsWith("gpt-5") && !modelId.startsWith("gpt-5-chat");
   const supportsPriorityProcessing = modelId.startsWith("gpt-4") || modelId.startsWith("gpt-5-mini") || modelId.startsWith("gpt-5") && !modelId.startsWith("gpt-5-nano") && !modelId.startsWith("gpt-5-chat") || modelId.startsWith("o3") || modelId.startsWith("o4-mini");
-  const isReasoningModel = !(modelId.startsWith("gpt-3") || modelId.startsWith("gpt-4") || modelId.startsWith("chatgpt-4o") || modelId.startsWith("gpt-5-chat"));
+  const isReasoningModel = modelId.startsWith("o1") || modelId.startsWith("o3") || modelId.startsWith("o4-mini") || modelId.startsWith("codex-mini") || modelId.startsWith("computer-use-preview") || modelId.startsWith("gpt-5") && !modelId.startsWith("gpt-5-chat");
   const supportsNonReasoningParameters = modelId.startsWith("gpt-5.1") || modelId.startsWith("gpt-5.2");
   const systemMessageMode = isReasoningModel ? "developer" : "system";
   return {
@@ -583,7 +583,26 @@ var openaiChatLanguageModelOptions = (0, import_provider_utils4.lazySchema)(
        * username or email address, in order to avoid sending us any identifying
        * information.
        */
-      safetyIdentifier: import_v43.z.string().optional()
+      safetyIdentifier: import_v43.z.string().optional(),
+      /**
+       * Override the system message mode for this model.
+       * - 'system': Use the 'system' role for system messages (default for most models)
+       * - 'developer': Use the 'developer' role for system messages (used by reasoning models)
+       * - 'remove': Remove system messages entirely
+       *
+       * If not specified, the mode is automatically determined based on the model.
+       */
+      systemMessageMode: import_v43.z.enum(["system", "developer", "remove"]).optional(),
+      /**
+       * Force treating this model as a reasoning model.
+       *
+       * This is useful for "stealth" reasoning models (e.g. via a custom baseURL)
+       * where the model ID is not recognized by the SDK's allowlist.
+       *
+       * When enabled, the SDK applies reasoning-model parameter compatibility rules
+       * and defaults `systemMessageMode` to `developer` unless overridden.
+       */
+      forceReasoning: import_v43.z.boolean().optional()
     })
   )
 );
@@ -678,7 +697,7 @@ var OpenAIChatLanguageModel = class {
     toolChoice,
     providerOptions
   }) {
-    var _a, _b, _c;
+    var _a, _b, _c, _d, _e;
     const warnings = [];
     const openaiOptions = (_a = await (0, import_provider_utils5.parseProviderOptions)({
       provider: "openai",
@@ -686,17 +705,18 @@ var OpenAIChatLanguageModel = class {
       schema: openaiChatLanguageModelOptions
     })) != null ? _a : {};
     const modelCapabilities = getOpenAILanguageModelCapabilities(this.modelId);
+    const isReasoningModel = (_b = openaiOptions.forceReasoning) != null ? _b : modelCapabilities.isReasoningModel;
     if (topK != null) {
       warnings.push({ type: "unsupported", feature: "topK" });
     }
     const { messages, warnings: messageWarnings } = convertToOpenAIChatMessages(
       {
         prompt,
-        systemMessageMode: modelCapabilities.systemMessageMode
+        systemMessageMode: (_c = openaiOptions.systemMessageMode) != null ? _c : isReasoningModel ? "developer" : modelCapabilities.systemMessageMode
       }
     );
     warnings.push(...messageWarnings);
-    const strictJsonSchema = (_b = openaiOptions.strictJsonSchema) != null ? _b : true;
+    const strictJsonSchema = (_d = openaiOptions.strictJsonSchema) != null ? _d : true;
     const baseArgs = {
       // model id:
       model: this.modelId,
@@ -717,7 +737,7 @@ var OpenAIChatLanguageModel = class {
         json_schema: {
           schema: responseFormat.schema,
           strict: strictJsonSchema,
-          name: (_c = responseFormat.name) != null ? _c : "response",
+          name: (_e = responseFormat.name) != null ? _e : "response",
           description: responseFormat.description
         }
       } : { type: "json_object" } : void 0,
@@ -738,7 +758,7 @@ var OpenAIChatLanguageModel = class {
       // messages:
       messages
     };
-    if (modelCapabilities.isReasoningModel) {
+    if (isReasoningModel) {
       if (openaiOptions.reasoningEffort !== "none" || !modelCapabilities.supportsNonReasoningParameters) {
         if (baseArgs.temperature != null) {
           baseArgs.temperature = void 0;
@@ -1733,11 +1753,13 @@ var modelMaxImagesPerCall = {
   "dall-e-3": 1,
   "dall-e-2": 10,
   "gpt-image-1": 10,
-  "gpt-image-1-mini": 10
+  "gpt-image-1-mini": 10,
+  "gpt-image-1.5": 10
 };
 var hasDefaultResponseFormat = /* @__PURE__ */ new Set([
   "gpt-image-1",
-  "gpt-image-1-mini"
+  "gpt-image-1-mini",
+  "gpt-image-1.5"
 ]);
 
 // src/image/openai-image-model.ts
@@ -1756,6 +1778,8 @@ var OpenAIImageModel = class {
   }
   async doGenerate({
     prompt,
+    files,
+    mask,
     n,
     size,
     aspectRatio,
@@ -1764,7 +1788,7 @@ var OpenAIImageModel = class {
     headers,
     abortSignal
   }) {
-    var _a, _b, _c, _d, _e, _f, _g;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k;
     const warnings = [];
     if (aspectRatio != null) {
       warnings.push({
@@ -1777,6 +1801,72 @@ var OpenAIImageModel = class {
       warnings.push({ type: "unsupported", feature: "seed" });
     }
     const currentDate = (_c = (_b = (_a = this.config._internal) == null ? void 0 : _a.currentDate) == null ? void 0 : _b.call(_a)) != null ? _c : /* @__PURE__ */ new Date();
+    if (files != null) {
+      const { value: response2, responseHeaders: responseHeaders2 } = await (0, import_provider_utils13.postFormDataToApi)({
+        url: this.config.url({
+          path: "/images/edits",
+          modelId: this.modelId
+        }),
+        headers: (0, import_provider_utils13.combineHeaders)(this.config.headers(), headers),
+        formData: (0, import_provider_utils13.convertToFormData)({
+          model: this.modelId,
+          prompt,
+          image: await Promise.all(
+            files.map(
+              (file) => file.type === "file" ? new Blob(
+                [
+                  file.data instanceof Uint8Array ? new Blob([file.data], {
+                    type: file.mediaType
+                  }) : new Blob([(0, import_provider_utils13.convertBase64ToUint8Array)(file.data)], {
+                    type: file.mediaType
+                  })
+                ],
+                { type: file.mediaType }
+              ) : (0, import_provider_utils13.downloadBlob)(file.url)
+            )
+          ),
+          mask: mask != null ? await fileToBlob(mask) : void 0,
+          n,
+          size,
+          ...(_d = providerOptions.openai) != null ? _d : {}
+        }),
+        failedResponseHandler: openaiFailedResponseHandler,
+        successfulResponseHandler: (0, import_provider_utils13.createJsonResponseHandler)(
+          openaiImageResponseSchema
+        ),
+        abortSignal,
+        fetch: this.config.fetch
+      });
+      return {
+        images: response2.data.map((item) => item.b64_json),
+        warnings,
+        usage: response2.usage != null ? {
+          inputTokens: (_e = response2.usage.input_tokens) != null ? _e : void 0,
+          outputTokens: (_f = response2.usage.output_tokens) != null ? _f : void 0,
+          totalTokens: (_g = response2.usage.total_tokens) != null ? _g : void 0
+        } : void 0,
+        response: {
+          timestamp: currentDate,
+          modelId: this.modelId,
+          headers: responseHeaders2
+        },
+        providerMetadata: {
+          openai: {
+            images: response2.data.map((item) => {
+              var _a2, _b2, _c2, _d2, _e2;
+              return {
+                ...item.revised_prompt ? { revisedPrompt: item.revised_prompt } : {},
+                created: (_a2 = response2.created) != null ? _a2 : void 0,
+                size: (_b2 = response2.size) != null ? _b2 : void 0,
+                quality: (_c2 = response2.quality) != null ? _c2 : void 0,
+                background: (_d2 = response2.background) != null ? _d2 : void 0,
+                outputFormat: (_e2 = response2.output_format) != null ? _e2 : void 0
+              };
+            })
+          }
+        }
+      };
+    }
     const { value: response, responseHeaders } = await (0, import_provider_utils13.postJsonToApi)({
       url: this.config.url({
         path: "/images/generations",
@@ -1788,7 +1878,7 @@ var OpenAIImageModel = class {
         prompt,
         n,
         size,
-        ...(_d = providerOptions.openai) != null ? _d : {},
+        ...(_h = providerOptions.openai) != null ? _h : {},
         ...!hasDefaultResponseFormat.has(this.modelId) ? { response_format: "b64_json" } : {}
       },
       failedResponseHandler: openaiFailedResponseHandler,
@@ -1802,9 +1892,9 @@ var OpenAIImageModel = class {
       images: response.data.map((item) => item.b64_json),
       warnings,
       usage: response.usage != null ? {
-        inputTokens: (_e = response.usage.input_tokens) != null ? _e : void 0,
-        outputTokens: (_f = response.usage.output_tokens) != null ? _f : void 0,
-        totalTokens: (_g = response.usage.total_tokens) != null ? _g : void 0
+        inputTokens: (_i = response.usage.input_tokens) != null ? _i : void 0,
+        outputTokens: (_j = response.usage.output_tokens) != null ? _j : void 0,
+        totalTokens: (_k = response.usage.total_tokens) != null ? _k : void 0
       } : void 0,
       response: {
         timestamp: currentDate,
@@ -1829,6 +1919,14 @@ var OpenAIImageModel = class {
     };
   }
 };
+async function fileToBlob(file) {
+  if (!file) return void 0;
+  if (file.type === "url") {
+    return (0, import_provider_utils13.downloadBlob)(file.url);
+  }
+  const data = file.data instanceof Uint8Array ? file.data : (0, import_provider_utils13.convertBase64ToUint8Array)(file.data);
+  return new Blob([data], { type: file.mediaType });
+}
 
 // src/transcription/openai-transcription-model.ts
 var import_provider_utils16 = require("@ai-sdk/provider-utils");
@@ -3578,7 +3676,26 @@ var openaiResponsesProviderOptionsSchema = (0, import_provider_utils24.lazySchem
        * Defaults to `undefined`.
        * @see https://platform.openai.com/docs/guides/safety-best-practices/end-user-ids
        */
-      user: import_v417.z.string().nullish()
+      user: import_v417.z.string().nullish(),
+      /**
+       * Override the system message mode for this model.
+       * - 'system': Use the 'system' role for system messages (default for most models)
+       * - 'developer': Use the 'developer' role for system messages (used by reasoning models)
+       * - 'remove': Remove system messages entirely
+       *
+       * If not specified, the mode is automatically determined based on the model.
+       */
+      systemMessageMode: import_v417.z.enum(["system", "developer", "remove"]).optional(),
+      /**
+       * Force treating this model as a reasoning model.
+       *
+       * This is useful for "stealth" reasoning models (e.g. via a custom baseURL)
+       * where the model ID is not recognized by the SDK's allowlist.
+       *
+       * When enabled, the SDK applies reasoning-model parameter compatibility rules
+       * and defaults `systemMessageMode` to `developer` unless overridden.
+       */
+      forceReasoning: import_v417.z.boolean().optional()
     })
   )
 );
@@ -4105,7 +4222,7 @@ var OpenAIResponsesLanguageModel = class {
     toolChoice,
     responseFormat
   }) {
-    var _a, _b, _c, _d;
+    var _a, _b, _c, _d, _e, _f;
     const warnings = [];
     const modelCapabilities = getOpenAILanguageModelCapabilities(this.modelId);
     if (topK != null) {
@@ -4128,6 +4245,7 @@ var OpenAIResponsesLanguageModel = class {
       providerOptions,
       schema: openaiResponsesProviderOptionsSchema
     });
+    const isReasoningModel = (_a = openaiOptions == null ? void 0 : openaiOptions.forceReasoning) != null ? _a : modelCapabilities.isReasoningModel;
     if ((openaiOptions == null ? void 0 : openaiOptions.conversation) && (openaiOptions == null ? void 0 : openaiOptions.previousResponseId)) {
       warnings.push({
         type: "unsupported",
@@ -4152,15 +4270,15 @@ var OpenAIResponsesLanguageModel = class {
     const { input, warnings: inputWarnings } = await convertToOpenAIResponsesInput({
       prompt,
       toolNameMapping,
-      systemMessageMode: modelCapabilities.systemMessageMode,
+      systemMessageMode: (_b = openaiOptions == null ? void 0 : openaiOptions.systemMessageMode) != null ? _b : isReasoningModel ? "developer" : modelCapabilities.systemMessageMode,
       fileIdPrefixes: this.config.fileIdPrefixes,
-      store: (_a = openaiOptions == null ? void 0 : openaiOptions.store) != null ? _a : true,
+      store: (_c = openaiOptions == null ? void 0 : openaiOptions.store) != null ? _c : true,
       hasLocalShellTool: hasOpenAITool("openai.local_shell"),
       hasShellTool: hasOpenAITool("openai.shell"),
       hasApplyPatchTool: hasOpenAITool("openai.apply_patch")
     });
     warnings.push(...inputWarnings);
-    const strictJsonSchema = (_b = openaiOptions == null ? void 0 : openaiOptions.strictJsonSchema) != null ? _b : true;
+    const strictJsonSchema = (_d = openaiOptions == null ? void 0 : openaiOptions.strictJsonSchema) != null ? _d : true;
     let include = openaiOptions == null ? void 0 : openaiOptions.include;
     function addInclude(key) {
       if (include == null) {
@@ -4176,9 +4294,9 @@ var OpenAIResponsesLanguageModel = class {
     if (topLogprobs) {
       addInclude("message.output_text.logprobs");
     }
-    const webSearchToolName = (_c = tools == null ? void 0 : tools.find(
+    const webSearchToolName = (_e = tools == null ? void 0 : tools.find(
       (tool) => tool.type === "provider" && (tool.id === "openai.web_search" || tool.id === "openai.web_search_preview")
-    )) == null ? void 0 : _c.name;
+    )) == null ? void 0 : _e.name;
     if (webSearchToolName) {
       addInclude("web_search_call.action.sources");
     }
@@ -4186,7 +4304,7 @@ var OpenAIResponsesLanguageModel = class {
       addInclude("code_interpreter_call.outputs");
     }
     const store = openaiOptions == null ? void 0 : openaiOptions.store;
-    if (store === false && modelCapabilities.isReasoningModel) {
+    if (store === false && isReasoningModel) {
       addInclude("reasoning.encrypted_content");
     }
     const baseArgs = {
@@ -4201,7 +4319,7 @@ var OpenAIResponsesLanguageModel = class {
             format: responseFormat.schema != null ? {
               type: "json_schema",
               strict: strictJsonSchema,
-              name: (_d = responseFormat.name) != null ? _d : "response",
+              name: (_f = responseFormat.name) != null ? _f : "response",
               description: responseFormat.description,
               schema: responseFormat.schema
             } : { type: "json_object" }
@@ -4228,7 +4346,7 @@ var OpenAIResponsesLanguageModel = class {
       top_logprobs: topLogprobs,
       truncation: openaiOptions == null ? void 0 : openaiOptions.truncation,
       // model-specific settings:
-      ...modelCapabilities.isReasoningModel && ((openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) != null || (openaiOptions == null ? void 0 : openaiOptions.reasoningSummary) != null) && {
+      ...isReasoningModel && ((openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) != null || (openaiOptions == null ? void 0 : openaiOptions.reasoningSummary) != null) && {
         reasoning: {
           ...(openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) != null && {
             effort: openaiOptions.reasoningEffort
@@ -4239,7 +4357,7 @@ var OpenAIResponsesLanguageModel = class {
         }
       }
     };
-    if (modelCapabilities.isReasoningModel) {
+    if (isReasoningModel) {
       if (!((openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) === "none" && modelCapabilities.supportsNonReasoningParameters)) {
         if (baseArgs.temperature != null) {
           baseArgs.temperature = void 0;