@ai-sdk/openai 3.0.83 → 3.0.85

package/dist/index.js

@@ -193,7 +193,7 @@ function isHttpErrorStatusCode(value) {
 
 // src/chat/convert-openai-chat-usage.ts
 function convertOpenAIChatUsage(usage) {
-  var _a, _b, _c, _d, _e, _f;
+  var _a, _b, _c, _d, _e, _f, _g, _h;
   if (usage == null) {
     return {
       inputTokens: {
@@ -213,13 +213,14 @@ function convertOpenAIChatUsage(usage) {
   const promptTokens = (_a = usage.prompt_tokens) != null ? _a : 0;
   const completionTokens = (_b = usage.completion_tokens) != null ? _b : 0;
   const cachedTokens = (_d = (_c = usage.prompt_tokens_details) == null ? void 0 : _c.cached_tokens) != null ? _d : 0;
-  const reasoningTokens = (_f = (_e = usage.completion_tokens_details) == null ? void 0 : _e.reasoning_tokens) != null ? _f : 0;
+  const cacheWriteTokens = (_f = (_e = usage.prompt_tokens_details) == null ? void 0 : _e.cache_write_tokens) != null ? _f : void 0;
+  const reasoningTokens = (_h = (_g = usage.completion_tokens_details) == null ? void 0 : _g.reasoning_tokens) != null ? _h : 0;
   return {
     inputTokens: {
       total: promptTokens,
-      noCache: promptTokens - cachedTokens,
+      noCache: promptTokens - cachedTokens - (cacheWriteTokens != null ? cacheWriteTokens : 0),
       cacheRead: cachedTokens,
-      cacheWrite: void 0
+      cacheWrite: cacheWriteTokens
     },
     outputTokens: {
       total: completionTokens,
@@ -236,23 +237,47 @@ var import_provider_utils2 = require("@ai-sdk/provider-utils");
 function serializeToolCallArguments(input) {
   return JSON.stringify(input === void 0 ? {} : input);
 }
+function getPromptCacheBreakpoint(providerOptions) {
+  var _a;
+  return (_a = providerOptions == null ? void 0 : providerOptions.openai) == null ? void 0 : _a.promptCacheBreakpoint;
+}
 function convertToOpenAIChatMessages({
   prompt,
   systemMessageMode = "system"
 }) {
-  var _a;
+  var _a, _b;
   const messages = [];
   const warnings = [];
-  for (const { role, content } of prompt) {
+  for (const { role, content, providerOptions } of prompt) {
     switch (role) {
       case "system": {
         switch (systemMessageMode) {
           case "system": {
-            messages.push({ role: "system", content });
+            const promptCacheBreakpoint = getPromptCacheBreakpoint(providerOptions);
+            messages.push({
+              role: "system",
+              content: promptCacheBreakpoint == null ? content : [
+                {
+                  type: "text",
+                  text: content,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ]
+            });
             break;
           }
           case "developer": {
-            messages.push({ role: "developer", content });
+            const promptCacheBreakpoint = getPromptCacheBreakpoint(providerOptions);
+            messages.push({
+              role: "developer",
+              content: promptCacheBreakpoint == null ? content : [
+                {
+                  type: "text",
+                  text: content,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ]
+            });
             break;
           }
           case "remove": {
@@ -272,19 +297,31 @@ function convertToOpenAIChatMessages({
         break;
       }
       case "user": {
-        if (content.length === 1 && content[0].type === "text") {
+        if (content.length === 1 && content[0].type === "text" && getPromptCacheBreakpoint(content[0].providerOptions) == null) {
           messages.push({ role: "user", content: content[0].text });
           break;
         }
         messages.push({
           role: "user",
           content: content.map((part, index) => {
-            var _a2, _b, _c;
+            var _a2, _b2, _c;
             switch (part.type) {
               case "text": {
-                return { type: "text", text: part.text };
+                const promptCacheBreakpoint = getPromptCacheBreakpoint(
+                  part.providerOptions
+                );
+                return {
+                  type: "text",
+                  text: part.text,
+                  ...promptCacheBreakpoint != null && {
+                    prompt_cache_breakpoint: promptCacheBreakpoint
+                  }
+                };
               }
               case "file": {
+                const promptCacheBreakpoint = getPromptCacheBreakpoint(
+                  part.providerOptions
+                );
                 if (part.mediaType.startsWith("image/")) {
                   const mediaType = part.mediaType === "image/*" ? "image/jpeg" : part.mediaType;
                   return {
@@ -292,7 +329,10 @@ function convertToOpenAIChatMessages({
                     image_url: {
                       url: part.data instanceof URL ? part.data.toString() : `data:${mediaType};base64,${(0, import_provider_utils2.convertToBase64)(part.data)}`,
                       // OpenAI specific extension: image detail
-                      detail: (_b = (_a2 = part.providerOptions) == null ? void 0 : _a2.openai) == null ? void 0 : _b.imageDetail
+                      detail: (_b2 = (_a2 = part.providerOptions) == null ? void 0 : _a2.openai) == null ? void 0 : _b2.imageDetail
+                    },
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
                     }
                   };
                 } else if (part.mediaType.startsWith("audio/")) {
@@ -308,6 +348,9 @@ function convertToOpenAIChatMessages({
                         input_audio: {
                           data: (0, import_provider_utils2.convertToBase64)(part.data),
                           format: "wav"
+                        },
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
                         }
                       };
                     }
@@ -318,6 +361,9 @@ function convertToOpenAIChatMessages({
                         input_audio: {
                           data: (0, import_provider_utils2.convertToBase64)(part.data),
                           format: "mp3"
+                        },
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
                         }
                       };
                     }
@@ -338,6 +384,9 @@ function convertToOpenAIChatMessages({
                     file: typeof part.data === "string" && part.data.startsWith("file-") ? { file_id: part.data } : {
                       filename: (_c = part.filename) != null ? _c : `part-${index}.pdf`,
                       file_data: `data:application/pdf;base64,${(0, import_provider_utils2.convertToBase64)(part.data)}`
+                    },
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
                     }
                   };
                 } else {
@@ -353,11 +402,24 @@ function convertToOpenAIChatMessages({
       }
       case "assistant": {
         let text = "";
+        const textParts = [];
+        let hasPromptCacheBreakpoint = false;
         const toolCalls = [];
         for (const part of content) {
           switch (part.type) {
             case "text": {
+              const promptCacheBreakpoint = getPromptCacheBreakpoint(
+                part.providerOptions
+              );
               text += part.text;
+              textParts.push({
+                type: "text",
+                text: part.text,
+                ...promptCacheBreakpoint != null && {
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              });
+              hasPromptCacheBreakpoint || (hasPromptCacheBreakpoint = promptCacheBreakpoint != null);
               break;
             }
             case "tool-call": {
@@ -375,7 +437,7 @@ function convertToOpenAIChatMessages({
         }
         messages.push({
           role: "assistant",
-          content: toolCalls.length > 0 ? text || null : text,
+          content: hasPromptCacheBreakpoint ? textParts : toolCalls.length > 0 ? text || null : text,
           tool_calls: toolCalls.length > 0 ? toolCalls : void 0
         });
         break;
@@ -386,6 +448,7 @@ function convertToOpenAIChatMessages({
             continue;
           }
           const output = toolResponse.output;
+          const promptCacheBreakpoint = (_a = output.type === "content" ? output.value.map((part) => getPromptCacheBreakpoint(part.providerOptions)).find((breakpoint) => breakpoint != null) : getPromptCacheBreakpoint(output.providerOptions)) != null ? _a : getPromptCacheBreakpoint(toolResponse.providerOptions);
           let contentValue;
           switch (output.type) {
             case "text":
@@ -393,7 +456,7 @@ function convertToOpenAIChatMessages({
               contentValue = output.value;
               break;
             case "execution-denied":
-              contentValue = (_a = output.reason) != null ? _a : "Tool call execution denied.";
+              contentValue = (_b = output.reason) != null ? _b : "Tool call execution denied.";
               break;
             case "content":
             case "json":
@@ -404,7 +467,13 @@ function convertToOpenAIChatMessages({
           messages.push({
             role: "tool",
             tool_call_id: toolResponse.toolCallId,
-            content: contentValue
+            content: promptCacheBreakpoint == null ? contentValue : [
+              {
+                type: "text",
+                text: contentValue,
+                prompt_cache_breakpoint: promptCacheBreakpoint
+              }
+            ]
           });
         }
         break;
@@ -507,7 +576,8 @@ var openaiChatResponseSchema = (0, import_provider_utils3.lazySchema)(
         completion_tokens: import_v42.z.number().nullish(),
         total_tokens: import_v42.z.number().nullish(),
         prompt_tokens_details: import_v42.z.object({
-          cached_tokens: import_v42.z.number().nullish()
+          cached_tokens: import_v42.z.number().nullish(),
+          cache_write_tokens: import_v42.z.number().nullish()
         }).nullish(),
         completion_tokens_details: import_v42.z.object({
           reasoning_tokens: import_v42.z.number().nullish(),
@@ -576,7 +646,8 @@ var openaiChatChunkSchema = (0, import_provider_utils3.lazySchema)(
           completion_tokens: import_v42.z.number().nullish(),
           total_tokens: import_v42.z.number().nullish(),
           prompt_tokens_details: import_v42.z.object({
-            cached_tokens: import_v42.z.number().nullish()
+            cached_tokens: import_v42.z.number().nullish(),
+            cache_write_tokens: import_v42.z.number().nullish()
           }).nullish(),
           completion_tokens_details: import_v42.z.object({
             reasoning_tokens: import_v42.z.number().nullish(),
@@ -625,7 +696,7 @@ var openaiLanguageModelChatOptions = (0, import_provider_utils4.lazySchema)(
       /**
        * Reasoning effort for reasoning models. Defaults to `medium`.
        */
-      reasoningEffort: import_v43.z.enum(["none", "minimal", "low", "medium", "high", "xhigh"]).optional(),
+      reasoningEffort: import_v43.z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]).optional(),
       /**
        * Maximum number of completion tokens to generate. Useful for reasoning models.
        */
@@ -669,11 +740,22 @@ var openaiLanguageModelChatOptions = (0, import_provider_utils4.lazySchema)(
        * Useful for improving cache hit rates and working around automatic caching issues.
        */
       promptCacheKey: import_v43.z.string().optional(),
+      /**
+       * Prompt cache behavior for GPT-5.6 and later models.
+       * `mode` controls whether OpenAI also places an implicit breakpoint.
+       * `ttl` sets the minimum cache lifetime and currently only supports 30 minutes.
+       */
+      promptCacheOptions: import_v43.z.object({
+        mode: import_v43.z.enum(["implicit", "explicit"]).optional(),
+        ttl: import_v43.z.literal("30m").optional()
+      }).optional(),
       /**
        * The retention policy for the prompt cache.
        * - 'in_memory': Default. Standard prompt caching behavior.
        * - '24h': Extended prompt caching that keeps cached prefixes active for up to 24 hours.
-       *          Currently only available for 5.1 series models.
+       *          Available for models before GPT-5.6 that support extended caching.
+       *
+       * @deprecated For GPT-5.6 and later models, use `promptCacheOptions.ttl`.
        *
        * @default 'in_memory'
        */
@@ -855,6 +937,7 @@ var OpenAIChatLanguageModel = class {
       reasoning_effort: openaiOptions.reasoningEffort,
       service_tier: openaiOptions.serviceTier,
       prompt_cache_key: openaiOptions.promptCacheKey,
+      prompt_cache_options: openaiOptions.promptCacheOptions,
       prompt_cache_retention: openaiOptions.promptCacheRetention,
       safety_identifier: openaiOptions.safetyIdentifier,
       // messages:
@@ -2833,7 +2916,7 @@ var import_provider_utils30 = require("@ai-sdk/provider-utils");
 
 // src/responses/convert-openai-responses-usage.ts
 function convertOpenAIResponsesUsage(usage) {
-  var _a, _b, _c, _d;
+  var _a, _b, _c, _d, _e, _f;
   if (usage == null) {
     return {
       inputTokens: {
@@ -2853,13 +2936,14 @@ function convertOpenAIResponsesUsage(usage) {
   const inputTokens = usage.input_tokens;
   const outputTokens = usage.output_tokens;
   const cachedTokens = (_b = (_a = usage.input_tokens_details) == null ? void 0 : _a.cached_tokens) != null ? _b : 0;
-  const reasoningTokens = (_d = (_c = usage.output_tokens_details) == null ? void 0 : _c.reasoning_tokens) != null ? _d : 0;
+  const cacheWriteTokens = (_d = (_c = usage.input_tokens_details) == null ? void 0 : _c.cache_write_tokens) != null ? _d : void 0;
+  const reasoningTokens = (_f = (_e = usage.output_tokens_details) == null ? void 0 : _e.reasoning_tokens) != null ? _f : 0;
   return {
     inputTokens: {
       total: inputTokens,
-      noCache: inputTokens - cachedTokens,
+      noCache: inputTokens - cachedTokens - (cacheWriteTokens != null ? cacheWriteTokens : 0),
       cacheRead: cachedTokens,
-      cacheWrite: void 0
+      cacheWrite: cacheWriteTokens
     },
     outputTokens: {
       total: outputTokens,
@@ -2877,6 +2961,10 @@ var import_v421 = require("zod/v4");
 function serializeToolCallArguments2(input) {
   return JSON.stringify(input === void 0 ? {} : input);
 }
+function getPromptCacheBreakpoint2(providerOptions, providerOptionsName) {
+  var _a;
+  return (_a = providerOptions == null ? void 0 : providerOptions[providerOptionsName]) == null ? void 0 : _a.promptCacheBreakpoint;
+}
 function isFileId(data, prefixes) {
   if (!prefixes) return false;
   return prefixes.some((prefix) => data.startsWith(prefix));
@@ -2900,16 +2988,42 @@ async function convertToOpenAIResponsesInput({
   let input = [];
   const warnings = [];
   const processedApprovalIds = /* @__PURE__ */ new Set();
-  for (const { role, content } of prompt) {
+  for (const { role, content, providerOptions } of prompt) {
     switch (role) {
       case "system": {
         switch (systemMessageMode) {
           case "system": {
-            input.push({ role: "system", content });
+            const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+              providerOptions,
+              providerOptionsName
+            );
+            input.push({
+              role: "system",
+              content: promptCacheBreakpoint == null ? content : [
+                {
+                  type: "input_text",
+                  text: content,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ]
+            });
             break;
           }
           case "developer": {
-            input.push({ role: "developer", content });
+            const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+              providerOptions,
+              providerOptionsName
+            );
+            input.push({
+              role: "developer",
+              content: promptCacheBreakpoint == null ? content : [
+                {
+                  type: "input_text",
+                  text: content,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ]
+            });
             break;
           }
           case "remove": {
@@ -2935,9 +3049,23 @@ async function convertToOpenAIResponsesInput({
             var _a2, _b2, _c2;
             switch (part.type) {
               case "text": {
-                return { type: "input_text", text: part.text };
+                const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                  part.providerOptions,
+                  providerOptionsName
+                );
+                return {
+                  type: "input_text",
+                  text: part.text,
+                  ...promptCacheBreakpoint != null && {
+                    prompt_cache_breakpoint: promptCacheBreakpoint
+                  }
+                };
               }
               case "file": {
+                const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                  part.providerOptions,
+                  providerOptionsName
+                );
                 const mediaType = part.mediaType === "image/*" ? "image/jpeg" : part.mediaType;
                 if (mediaType.startsWith("image/")) {
                   return {
@@ -2945,13 +3073,19 @@ async function convertToOpenAIResponsesInput({
                     ...part.data instanceof URL ? { image_url: part.data.toString() } : typeof part.data === "string" && isFileId(part.data, fileIdPrefixes) ? { file_id: part.data } : {
                       image_url: `data:${mediaType};base64,${(0, import_provider_utils26.convertToBase64)(part.data)}`
                     },
-                    detail: (_b2 = (_a2 = part.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail
+                    detail: (_b2 = (_a2 = part.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail,
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
+                    }
                   };
                 }
                 if (part.data instanceof URL) {
                   return {
                     type: "input_file",
-                    file_url: part.data.toString()
+                    file_url: part.data.toString(),
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
+                    }
                   };
                 }
                 if (mediaType !== "application/pdf" && !passThroughUnsupportedFiles) {
@@ -2964,6 +3098,9 @@ async function convertToOpenAIResponsesInput({
                   ...typeof part.data === "string" && isFileId(part.data, fileIdPrefixes) ? { file_id: part.data } : {
                     filename: (_c2 = part.filename) != null ? _c2 : mediaType === "application/pdf" ? `part-${index}.pdf` : `part-${index}`,
                     file_data: `data:${mediaType};base64,${(0, import_provider_utils26.convertToBase64)(part.data)}`
+                  },
+                  ...promptCacheBreakpoint != null && {
+                    prompt_cache_breakpoint: promptCacheBreakpoint
                   }
                 };
               }
@@ -3176,12 +3313,12 @@ async function convertToOpenAIResponsesInput({
               break;
             }
             case "reasoning": {
-              const providerOptions = await (0, import_provider_utils26.parseProviderOptions)({
+              const providerOptions2 = await (0, import_provider_utils26.parseProviderOptions)({
                 provider: providerOptionsName,
                 providerOptions: part.providerOptions,
                 schema: openaiResponsesReasoningProviderOptionsSchema
               });
-              const reasoningId = providerOptions == null ? void 0 : providerOptions.itemId;
+              const reasoningId = providerOptions2 == null ? void 0 : providerOptions2.itemId;
               if ((hasConversation || hasPreviousResponseId) && reasoningId != null) {
                 break;
               }
@@ -3213,19 +3350,19 @@ async function convertToOpenAIResponsesInput({
                     reasoningMessages[reasoningId] = {
                       type: "reasoning",
                       id: reasoningId,
-                      encrypted_content: providerOptions == null ? void 0 : providerOptions.reasoningEncryptedContent,
+                      encrypted_content: providerOptions2 == null ? void 0 : providerOptions2.reasoningEncryptedContent,
                       summary: summaryParts
                     };
                     input.push(reasoningMessages[reasoningId]);
                   } else {
                     reasoningMessage.summary.push(...summaryParts);
-                    if ((providerOptions == null ? void 0 : providerOptions.reasoningEncryptedContent) != null) {
-                      reasoningMessage.encrypted_content = providerOptions.reasoningEncryptedContent;
+                    if ((providerOptions2 == null ? void 0 : providerOptions2.reasoningEncryptedContent) != null) {
+                      reasoningMessage.encrypted_content = providerOptions2.reasoningEncryptedContent;
                     }
                   }
                 }
               } else {
-                const encryptedContent = providerOptions == null ? void 0 : providerOptions.reasoningEncryptedContent;
+                const encryptedContent = providerOptions2 == null ? void 0 : providerOptions2.reasoningEncryptedContent;
                 if (encryptedContent != null) {
                   const summaryParts = [];
                   if (part.text.length > 0) {
@@ -3358,31 +3495,53 @@ async function convertToOpenAIResponsesInput({
               case "content":
                 outputValue = output.value.map((item) => {
                   var _a2, _b2, _c2, _d2, _e2;
+                  const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                    item.providerOptions,
+                    providerOptionsName
+                  );
                   switch (item.type) {
                     case "text":
-                      return { type: "input_text", text: item.text };
+                      return {
+                        type: "input_text",
+                        text: item.text,
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
+                        }
+                      };
                     case "image-data":
                       return {
                         type: "input_image",
                         image_url: `data:${item.mediaType};base64,${item.data}`,
-                        detail: (_b2 = (_a2 = item.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail
+                        detail: (_b2 = (_a2 = item.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail,
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
+                        }
                       };
                     case "image-url":
                       return {
                         type: "input_image",
                         image_url: item.url,
-                        detail: (_d2 = (_c2 = item.providerOptions) == null ? void 0 : _c2[providerOptionsName]) == null ? void 0 : _d2.imageDetail
+                        detail: (_d2 = (_c2 = item.providerOptions) == null ? void 0 : _c2[providerOptionsName]) == null ? void 0 : _d2.imageDetail,
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
+                        }
                       };
                     case "file-data":
                       return {
                         type: "input_file",
                         filename: (_e2 = item.filename) != null ? _e2 : "data",
-                        file_data: `data:${item.mediaType};base64,${item.data}`
+                        file_data: `data:${item.mediaType};base64,${item.data}`,
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
+                        }
                       };
                     case "file-url":
                       return {
                         type: "input_file",
-                        file_url: item.url
+                        file_url: item.url,
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
+                        }
                       };
                     default:
                       warnings.push({
@@ -3419,35 +3578,57 @@ async function convertToOpenAIResponsesInput({
             case "content":
               contentValue = output.value.map((item) => {
                 var _a2, _b2, _c2, _d2, _e2;
+                const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                  item.providerOptions,
+                  providerOptionsName
+                );
                 switch (item.type) {
                   case "text": {
-                    return { type: "input_text", text: item.text };
+                    return {
+                      type: "input_text",
+                      text: item.text,
+                      ...promptCacheBreakpoint != null && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
+                    };
                   }
                   case "image-data": {
                     return {
                       type: "input_image",
                       image_url: `data:${item.mediaType};base64,${item.data}`,
-                      detail: (_b2 = (_a2 = item.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail
+                      detail: (_b2 = (_a2 = item.providerOptions) == null ? void 0 : _a2[providerOptionsName]) == null ? void 0 : _b2.imageDetail,
+                      ...promptCacheBreakpoint != null && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     };
                   }
                   case "image-url": {
                     return {
                       type: "input_image",
                       image_url: item.url,
-                      detail: (_d2 = (_c2 = item.providerOptions) == null ? void 0 : _c2[providerOptionsName]) == null ? void 0 : _d2.imageDetail
+                      detail: (_d2 = (_c2 = item.providerOptions) == null ? void 0 : _c2[providerOptionsName]) == null ? void 0 : _d2.imageDetail,
+                      ...promptCacheBreakpoint != null && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     };
                   }
                   case "file-data": {
                     return {
                       type: "input_file",
                       filename: (_e2 = item.filename) != null ? _e2 : "data",
-                      file_data: `data:${item.mediaType};base64,${item.data}`
+                      file_data: `data:${item.mediaType};base64,${item.data}`,
+                      ...promptCacheBreakpoint != null && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     };
                   }
                   case "file-url": {
                     return {
                       type: "input_file",
-                      file_url: item.url
+                      file_url: item.url,
+                      ...promptCacheBreakpoint != null && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     };
                   }
                   default: {
@@ -3569,6 +3750,7 @@ var openaiResponsesChunkSchema = (0, import_provider_utils27.lazySchema)(
             input_tokens: import_v422.z.number(),
             input_tokens_details: import_v422.z.object({
               cached_tokens: import_v422.z.number().nullish(),
+              cache_write_tokens: import_v422.z.number().nullish(),
               orchestration_input_tokens: import_v422.z.number().nullish(),
               orchestration_input_cached_tokens: import_v422.z.number().nullish()
             }).nullish(),
@@ -3578,6 +3760,9 @@ var openaiResponsesChunkSchema = (0, import_provider_utils27.lazySchema)(
               orchestration_output_tokens: import_v422.z.number().nullish()
             }).nullish()
           }),
+          reasoning: import_v422.z.object({
+            context: import_v422.z.string().nullish()
+          }).nullish(),
           service_tier: import_v422.z.string().nullish()
         })
       }),
@@ -3594,6 +3779,7 @@ var openaiResponsesChunkSchema = (0, import_provider_utils27.lazySchema)(
             input_tokens: import_v422.z.number(),
             input_tokens_details: import_v422.z.object({
               cached_tokens: import_v422.z.number().nullish(),
+              cache_write_tokens: import_v422.z.number().nullish(),
               orchestration_input_tokens: import_v422.z.number().nullish(),
               orchestration_input_cached_tokens: import_v422.z.number().nullish()
             }).nullish(),
@@ -3603,6 +3789,9 @@ var openaiResponsesChunkSchema = (0, import_provider_utils27.lazySchema)(
               orchestration_output_tokens: import_v422.z.number().nullish()
             }).nullish()
           }).nullish(),
+          reasoning: import_v422.z.object({
+            context: import_v422.z.string().nullish()
+          }).nullish(),
           service_tier: import_v422.z.string().nullish()
         })
       }),
@@ -4374,11 +4563,15 @@ var openaiResponsesResponseSchema = (0, import_provider_utils27.lazySchema)(
         ])
       ).optional(),
       service_tier: import_v422.z.string().nullish(),
+      reasoning: import_v422.z.object({
+        context: import_v422.z.string().nullish()
+      }).nullish(),
       incomplete_details: import_v422.z.object({ reason: import_v422.z.string() }).nullish(),
       usage: import_v422.z.object({
         input_tokens: import_v422.z.number(),
         input_tokens_details: import_v422.z.object({
           cached_tokens: import_v422.z.number().nullish(),
+          cache_write_tokens: import_v422.z.number().nullish(),
           orchestration_input_tokens: import_v422.z.number().nullish(),
           orchestration_input_cached_tokens: import_v422.z.number().nullish()
         }).nullish(),
@@ -4533,11 +4726,22 @@ var openaiLanguageModelResponsesOptionsSchema = (0, import_provider_utils28.lazy
        * Sets a cache key to tie this prompt to cached prefixes for better caching performance.
        */
       promptCacheKey: import_v423.z.string().nullish(),
+      /**
+       * Prompt cache behavior for GPT-5.6 and later models.
+       * `mode` controls whether OpenAI also places an implicit breakpoint.
+       * `ttl` sets the minimum cache lifetime and currently only supports 30 minutes.
+       */
+      promptCacheOptions: import_v423.z.object({
+        mode: import_v423.z.enum(["implicit", "explicit"]).optional(),
+        ttl: import_v423.z.literal("30m").optional()
+      }).optional(),
       /**
        * The retention policy for the prompt cache.
        * - 'in_memory': Default. Standard prompt caching behavior.
        * - '24h': Extended prompt caching that keeps cached prefixes active for up to 24 hours.
-       *          Currently only available for 5.1 series models.
+       *          Available for models before GPT-5.6 that support extended caching.
+       *
+       * @deprecated For GPT-5.6 and later models, use `promptCacheOptions.ttl`.
        *
        * @default 'in_memory'
        */
@@ -4545,14 +4749,21 @@ var openaiLanguageModelResponsesOptionsSchema = (0, import_provider_utils28.lazy
       /**
        * Reasoning effort for reasoning models. Defaults to `medium`. If you use
        * `providerOptions` to set the `reasoningEffort` option, this model setting will be ignored.
-       * Valid values: 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
-       *
-       * The 'none' type for `reasoningEffort` is only available for OpenAI's GPT-5.1
-       * models. Also, the 'xhigh' type for `reasoningEffort` is only available for
-       * OpenAI's GPT-5.1-Codex-Max model. Setting `reasoningEffort` to 'none' or 'xhigh' with unsupported models will result in
-       * an error.
+       * GPT-5.6 supports 'none' | 'low' | 'medium' | 'high' | 'xhigh' | 'max'.
+       * Supported values vary by model.
        */
       reasoningEffort: import_v423.z.string().nullish(),
+      /**
+       * Controls how much model work GPT-5.6 performs before returning a final answer.
+       * `standard` is the default. `pro` increases quality, latency, and token usage.
+       */
+      reasoningMode: import_v423.z.enum(["standard", "pro"]).optional(),
+      /**
+       * Controls which available reasoning items GPT-5.6 can use.
+       * `auto` uses the model default, `current_turn` excludes reasoning from earlier
+       * turns, and `all_turns` makes compatible earlier reasoning available.
+       */
+      reasoningContext: import_v423.z.enum(["auto", "current_turn", "all_turns"]).optional(),
       /**
        * Controls reasoning summary output from the model.
        * Set to "auto" to automatically receive the richest level available,
@@ -5154,18 +5365,25 @@ var OpenAIResponsesLanguageModel = class {
       service_tier: openaiOptions == null ? void 0 : openaiOptions.serviceTier,
       include,
       prompt_cache_key: openaiOptions == null ? void 0 : openaiOptions.promptCacheKey,
+      prompt_cache_options: openaiOptions == null ? void 0 : openaiOptions.promptCacheOptions,
       prompt_cache_retention: openaiOptions == null ? void 0 : openaiOptions.promptCacheRetention,
       safety_identifier: openaiOptions == null ? void 0 : openaiOptions.safetyIdentifier,
       top_logprobs: topLogprobs,
       truncation: openaiOptions == null ? void 0 : openaiOptions.truncation,
       // model-specific settings:
-      ...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 || (openaiOptions == null ? void 0 : openaiOptions.reasoningMode) != null || (openaiOptions == null ? void 0 : openaiOptions.reasoningContext) != null) && {
         reasoning: {
           ...(openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) != null && {
             effort: openaiOptions.reasoningEffort
           },
           ...(openaiOptions == null ? void 0 : openaiOptions.reasoningSummary) != null && {
             summary: openaiOptions.reasoningSummary
+          },
+          ...(openaiOptions == null ? void 0 : openaiOptions.reasoningMode) != null && {
+            mode: openaiOptions.reasoningMode
+          },
+          ...(openaiOptions == null ? void 0 : openaiOptions.reasoningContext) != null && {
+            context: openaiOptions.reasoningContext
           }
         }
       }
@@ -5204,6 +5422,20 @@ var OpenAIResponsesLanguageModel = class {
           details: "reasoningSummary is not supported for non-reasoning models"
         });
       }
+      if ((openaiOptions == null ? void 0 : openaiOptions.reasoningMode) != null) {
+        warnings.push({
+          type: "unsupported",
+          feature: "reasoningMode",
+          details: "reasoningMode is not supported for non-reasoning models"
+        });
+      }
+      if ((openaiOptions == null ? void 0 : openaiOptions.reasoningContext) != null) {
+        warnings.push({
+          type: "unsupported",
+          feature: "reasoningContext",
+          details: "reasoningContext is not supported for non-reasoning models"
+        });
+      }
     }
     if ((openaiOptions == null ? void 0 : openaiOptions.serviceTier) === "flex" && !modelCapabilities.supportsFlexProcessing) {
       warnings.push({
@@ -5240,7 +5472,7 @@ var OpenAIResponsesLanguageModel = class {
     };
   }
   async doGenerate(options) {
-    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q, _r, _s, _t, _u, _v, _w, _x, _y, _z, _A, _B;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q, _r, _s, _t, _u, _v, _w, _x, _y, _z, _A, _B, _C;
     const {
       args: body,
       warnings,
@@ -5689,7 +5921,8 @@ var OpenAIResponsesLanguageModel = class {
       [providerOptionsName]: {
         responseId: response.id,
         ...logprobs.length > 0 ? { logprobs } : {},
-        ...typeof response.service_tier === "string" ? { serviceTier: response.service_tier } : {}
+        ...typeof response.service_tier === "string" ? { serviceTier: response.service_tier } : {},
+        ...((_z = response.reasoning) == null ? void 0 : _z.context) != null ? { reasoningContext: response.reasoning.context } : {}
       }
     };
     const usage = response.usage;
@@ -5697,10 +5930,10 @@ var OpenAIResponsesLanguageModel = class {
       content,
       finishReason: {
         unified: mapOpenAIResponseFinishReason({
-          finishReason: (_z = response.incomplete_details) == null ? void 0 : _z.reason,
+          finishReason: (_A = response.incomplete_details) == null ? void 0 : _A.reason,
           hasFunctionCall
         }),
-        raw: (_B = (_A = response.incomplete_details) == null ? void 0 : _A.reason) != null ? _B : void 0
+        raw: (_C = (_B = response.incomplete_details) == null ? void 0 : _B.reason) != null ? _C : void 0
       },
       usage: convertOpenAIResponsesUsage(usage),
       request: { body },
@@ -5767,6 +6000,7 @@ var OpenAIResponsesLanguageModel = class {
     let hasFunctionCall = false;
     const activeReasoning = {};
     let serviceTier;
+    let reasoningContext;
     const hostedToolSearchCallIds = [];
     let encounteredStreamError = false;
     const result = {
@@ -5776,7 +6010,7 @@ var OpenAIResponsesLanguageModel = class {
             controller.enqueue({ type: "stream-start", warnings });
           },
           transform(chunk, controller) {
-            var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q, _r, _s, _t, _u, _v, _w, _x, _y, _z, _A, _B, _C, _D, _E, _F, _G, _H, _I, _J, _K, _L;
+            var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q, _r, _s, _t, _u, _v, _w, _x, _y, _z, _A, _B, _C, _D, _E, _F, _G, _H, _I, _J, _K, _L, _M, _N;
             if (options.includeRawChunks) {
               controller.enqueue({ type: "raw", rawValue: chunk.rawValue });
             }
@@ -6518,8 +6752,11 @@ var OpenAIResponsesLanguageModel = class {
               if (typeof value.response.service_tier === "string") {
                 serviceTier = value.response.service_tier;
               }
+              if (((_y = value.response.reasoning) == null ? void 0 : _y.context) != null) {
+                reasoningContext = value.response.reasoning.context;
+              }
             } else if (isResponseFailedChunk(value)) {
-              const incompleteReason = (_y = value.response.incomplete_details) == null ? void 0 : _y.reason;
+              const incompleteReason = (_z = value.response.incomplete_details) == null ? void 0 : _z.reason;
               finishReason = {
                 unified: incompleteReason ? mapOpenAIResponseFinishReason({
                   finishReason: incompleteReason,
@@ -6527,7 +6764,10 @@ var OpenAIResponsesLanguageModel = class {
                 }) : "error",
                 raw: incompleteReason != null ? incompleteReason : "error"
               };
-              usage = (_z = value.response.usage) != null ? _z : void 0;
+              usage = (_A = value.response.usage) != null ? _A : void 0;
+              if (((_B = value.response.reasoning) == null ? void 0 : _B.context) != null) {
+                reasoningContext = value.response.reasoning.context;
+              }
               if (!encounteredStreamError && value.response.error != null) {
                 encounteredStreamError = true;
                 controller.enqueue({
@@ -6549,7 +6789,7 @@ var OpenAIResponsesLanguageModel = class {
                 controller.enqueue({
                   type: "source",
                   sourceType: "url",
-                  id: (_C = (_B = (_A = self.config).generateId) == null ? void 0 : _B.call(_A)) != null ? _C : (0, import_provider_utils30.generateId)(),
+                  id: (_E = (_D = (_C = self.config).generateId) == null ? void 0 : _D.call(_C)) != null ? _E : (0, import_provider_utils30.generateId)(),
                   url: value.annotation.url,
                   title: value.annotation.title
                 });
@@ -6557,7 +6797,7 @@ var OpenAIResponsesLanguageModel = class {
                 controller.enqueue({
                   type: "source",
                   sourceType: "document",
-                  id: (_F = (_E = (_D = self.config).generateId) == null ? void 0 : _E.call(_D)) != null ? _F : (0, import_provider_utils30.generateId)(),
+                  id: (_H = (_G = (_F = self.config).generateId) == null ? void 0 : _G.call(_F)) != null ? _H : (0, import_provider_utils30.generateId)(),
                   mediaType: "text/plain",
                   title: value.annotation.filename,
                   filename: value.annotation.filename,
@@ -6573,7 +6813,7 @@ var OpenAIResponsesLanguageModel = class {
                 controller.enqueue({
                   type: "source",
                   sourceType: "document",
-                  id: (_I = (_H = (_G = self.config).generateId) == null ? void 0 : _H.call(_G)) != null ? _I : (0, import_provider_utils30.generateId)(),
+                  id: (_K = (_J = (_I = self.config).generateId) == null ? void 0 : _J.call(_I)) != null ? _K : (0, import_provider_utils30.generateId)(),
                   mediaType: "text/plain",
                   title: value.annotation.filename,
                   filename: value.annotation.filename,
@@ -6589,7 +6829,7 @@ var OpenAIResponsesLanguageModel = class {
                 controller.enqueue({
                   type: "source",
                   sourceType: "document",
-                  id: (_L = (_K = (_J = self.config).generateId) == null ? void 0 : _K.call(_J)) != null ? _L : (0, import_provider_utils30.generateId)(),
+                  id: (_N = (_M = (_L = self.config).generateId) == null ? void 0 : _M.call(_L)) != null ? _N : (0, import_provider_utils30.generateId)(),
                   mediaType: "application/octet-stream",
                   title: value.annotation.file_id,
                   filename: value.annotation.file_id,
@@ -6613,7 +6853,8 @@ var OpenAIResponsesLanguageModel = class {
               [providerOptionsName]: {
                 responseId,
                 ...logprobs.length > 0 ? { logprobs } : {},
-                ...serviceTier !== void 0 ? { serviceTier } : {}
+                ...serviceTier !== void 0 ? { serviceTier } : {},
+                ...reasoningContext !== void 0 ? { reasoningContext } : {}
               }
             };
             controller.enqueue({
@@ -7089,7 +7330,7 @@ var OpenAITranscriptionModel = class {
 };
 
 // src/version.ts
-var VERSION = true ? "3.0.83" : "0.0.0-test";
+var VERSION = true ? "3.0.85" : "0.0.0-test";
 
 // src/openai-provider.ts
 function createOpenAI(options = {}) {