@ai-sdk/openai 2.0.109 → 2.0.111

package/dist/internal/index.mjs

@@ -36,7 +36,7 @@ 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") && !modelId.startsWith("gpt-5-nano") && !modelId.startsWith("gpt-5-chat") && !modelId.startsWith("gpt-5.4-nano") || 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 supportsNonReasoningParameters = modelId.startsWith("gpt-5.1") || modelId.startsWith("gpt-5.2") || modelId.startsWith("gpt-5.3") || modelId.startsWith("gpt-5.4");
+  const supportsNonReasoningParameters = modelId.startsWith("gpt-5.1") || modelId.startsWith("gpt-5.2") || modelId.startsWith("gpt-5.3") || modelId.startsWith("gpt-5.4") || modelId.startsWith("gpt-5.6");
   const systemMessageMode = isReasoningModel ? "developer" : "system";
   return {
     supportsFlexProcessing,
@@ -52,22 +52,46 @@ import {
   UnsupportedFunctionalityError
 } from "@ai-sdk/provider";
 import { convertToBase64 } from "@ai-sdk/provider-utils";
+function getPromptCacheBreakpoint(providerOptions) {
+  var _a;
+  return (_a = providerOptions == null ? void 0 : providerOptions.openai) == null ? void 0 : _a.promptCacheBreakpoint;
+}
 function convertToOpenAIChatMessages({
   prompt,
   systemMessageMode = "system"
 }) {
   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": {
@@ -87,7 +111,7 @@ 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;
         }
@@ -97,9 +121,21 @@ function convertToOpenAIChatMessages({
             var _a, _b, _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 {
@@ -108,6 +144,9 @@ function convertToOpenAIChatMessages({
                       url: part.data instanceof URL ? part.data.toString() : `data:${mediaType};base64,${convertToBase64(part.data)}`,
                       // OpenAI specific extension: image detail
                       detail: (_b = (_a = part.providerOptions) == null ? void 0 : _a.openai) == null ? void 0 : _b.imageDetail
+                    },
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
                     }
                   };
                 } else if (part.mediaType.startsWith("audio/")) {
@@ -123,6 +162,9 @@ function convertToOpenAIChatMessages({
                         input_audio: {
                           data: convertToBase64(part.data),
                           format: "wav"
+                        },
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
                         }
                       };
                     }
@@ -133,6 +175,9 @@ function convertToOpenAIChatMessages({
                         input_audio: {
                           data: convertToBase64(part.data),
                           format: "mp3"
+                        },
+                        ...promptCacheBreakpoint != null && {
+                          prompt_cache_breakpoint: promptCacheBreakpoint
                         }
                       };
                     }
@@ -153,6 +198,9 @@ function convertToOpenAIChatMessages({
                     file: typeof part.data === "string" && part.data.startsWith("file-") ? { file_id: part.data } : {
                       filename: (_c = part.filename) != null ? _c : `part-${index}.pdf`,
                       file_data: `data:application/pdf;base64,${convertToBase64(part.data)}`
+                    },
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
                     }
                   };
                 } else {
@@ -168,11 +216,24 @@ 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": {
@@ -190,7 +251,7 @@ function convertToOpenAIChatMessages({
         }
         messages.push({
           role: "assistant",
-          content: text,
+          content: hasPromptCacheBreakpoint ? textParts : text,
           tool_calls: toolCalls.length > 0 ? toolCalls : void 0
         });
         break;
@@ -198,6 +259,9 @@ function convertToOpenAIChatMessages({
       case "tool": {
         for (const toolResponse of content) {
           const output = toolResponse.output;
+          const promptCacheBreakpoint = getPromptCacheBreakpoint(
+            toolResponse.providerOptions
+          );
           let contentValue;
           switch (output.type) {
             case "text":
@@ -213,7 +277,13 @@ 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;
@@ -319,7 +389,8 @@ var openaiChatResponseSchema = lazyValidator(
         completion_tokens: z2.number().nullish(),
         total_tokens: z2.number().nullish(),
         prompt_tokens_details: z2.object({
-          cached_tokens: z2.number().nullish()
+          cached_tokens: z2.number().nullish(),
+          cache_write_tokens: z2.number().nullish()
         }).nullish(),
         completion_tokens_details: z2.object({
           reasoning_tokens: z2.number().nullish(),
@@ -388,7 +459,8 @@ var openaiChatChunkSchema = lazyValidator(
           completion_tokens: z2.number().nullish(),
           total_tokens: z2.number().nullish(),
           prompt_tokens_details: z2.object({
-            cached_tokens: z2.number().nullish()
+            cached_tokens: z2.number().nullish(),
+            cache_write_tokens: z2.number().nullish()
           }).nullish(),
           completion_tokens_details: z2.object({
             reasoning_tokens: z2.number().nullish(),
@@ -440,7 +512,7 @@ var openaiChatLanguageModelOptions = lazyValidator2(
       /**
        * Reasoning effort for reasoning models. Defaults to `medium`.
        */
-      reasoningEffort: z3.enum(["none", "minimal", "low", "medium", "high", "xhigh"]).optional(),
+      reasoningEffort: z3.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]).optional(),
       /**
        * Maximum number of completion tokens to generate. Useful for reasoning models.
        */
@@ -490,11 +562,22 @@ var openaiChatLanguageModelOptions = lazyValidator2(
        * Useful for improving cache hit rates and working around automatic caching issues.
        */
       promptCacheKey: z3.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: z3.object({
+        mode: z3.enum(["implicit", "explicit"]).optional(),
+        ttl: z3.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'
        */
@@ -668,6 +751,7 @@ 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:
@@ -833,6 +917,11 @@ var OpenAIChatLanguageModel = class {
     if ((completionTokenDetails == null ? void 0 : completionTokenDetails.rejected_prediction_tokens) != null) {
       providerMetadata.openai.rejectedPredictionTokens = completionTokenDetails == null ? void 0 : completionTokenDetails.rejected_prediction_tokens;
     }
+    if ((promptTokenDetails == null ? void 0 : promptTokenDetails.cache_write_tokens) != null) {
+      providerMetadata.openai.usage = {
+        cacheWriteTokens: promptTokenDetails.cache_write_tokens
+      };
+    }
     if (((_f = choice.logprobs) == null ? void 0 : _f.content) != null) {
       providerMetadata.openai.logprobs = choice.logprobs.content;
     }
@@ -896,7 +985,7 @@ var OpenAIChatLanguageModel = 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;
+            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;
             if (options.includeRawChunks) {
               controller.enqueue({ type: "raw", rawValue: chunk.rawValue });
             }
@@ -927,18 +1016,23 @@ var OpenAIChatLanguageModel = class {
               usage.totalTokens = (_c = value.usage.total_tokens) != null ? _c : void 0;
               usage.reasoningTokens = (_e = (_d = value.usage.completion_tokens_details) == null ? void 0 : _d.reasoning_tokens) != null ? _e : void 0;
               usage.cachedInputTokens = (_g = (_f = value.usage.prompt_tokens_details) == null ? void 0 : _f.cached_tokens) != null ? _g : void 0;
-              if (((_h = value.usage.completion_tokens_details) == null ? void 0 : _h.accepted_prediction_tokens) != null) {
-                providerMetadata.openai.acceptedPredictionTokens = (_i = value.usage.completion_tokens_details) == null ? void 0 : _i.accepted_prediction_tokens;
+              if (((_h = value.usage.prompt_tokens_details) == null ? void 0 : _h.cache_write_tokens) != null) {
+                providerMetadata.openai.usage = {
+                  cacheWriteTokens: value.usage.prompt_tokens_details.cache_write_tokens
+                };
+              }
+              if (((_i = value.usage.completion_tokens_details) == null ? void 0 : _i.accepted_prediction_tokens) != null) {
+                providerMetadata.openai.acceptedPredictionTokens = (_j = value.usage.completion_tokens_details) == null ? void 0 : _j.accepted_prediction_tokens;
               }
-              if (((_j = value.usage.completion_tokens_details) == null ? void 0 : _j.rejected_prediction_tokens) != null) {
-                providerMetadata.openai.rejectedPredictionTokens = (_k = value.usage.completion_tokens_details) == null ? void 0 : _k.rejected_prediction_tokens;
+              if (((_k = value.usage.completion_tokens_details) == null ? void 0 : _k.rejected_prediction_tokens) != null) {
+                providerMetadata.openai.rejectedPredictionTokens = (_l = value.usage.completion_tokens_details) == null ? void 0 : _l.rejected_prediction_tokens;
               }
             }
             const choice = value.choices[0];
             if ((choice == null ? void 0 : choice.finish_reason) != null) {
               finishReason = mapOpenAIFinishReason(choice.finish_reason);
             }
-            if (((_l = choice == null ? void 0 : choice.logprobs) == null ? void 0 : _l.content) != null) {
+            if (((_m = choice == null ? void 0 : choice.logprobs) == null ? void 0 : _m.content) != null) {
               providerMetadata.openai.logprobs = choice.logprobs.content;
             }
             if ((choice == null ? void 0 : choice.delta) == null) {
@@ -972,7 +1066,7 @@ var OpenAIChatLanguageModel = class {
                       message: `Expected 'id' to be a string.`
                     });
                   }
-                  if (((_m = toolCallDelta.function) == null ? void 0 : _m.name) == null) {
+                  if (((_n = toolCallDelta.function) == null ? void 0 : _n.name) == null) {
                     throw new InvalidResponseDataError({
                       data: toolCallDelta,
                       message: `Expected 'function.name' to be a string.`
@@ -988,12 +1082,12 @@ var OpenAIChatLanguageModel = class {
                     type: "function",
                     function: {
                       name: toolCallDelta.function.name,
-                      arguments: (_n = toolCallDelta.function.arguments) != null ? _n : ""
+                      arguments: (_o = toolCallDelta.function.arguments) != null ? _o : ""
                     },
                     hasFinished: false
                   };
                   const toolCall2 = toolCalls[index];
-                  if (((_o = toolCall2.function) == null ? void 0 : _o.name) != null && ((_p = toolCall2.function) == null ? void 0 : _p.arguments) != null) {
+                  if (((_p = toolCall2.function) == null ? void 0 : _p.name) != null && ((_q = toolCall2.function) == null ? void 0 : _q.arguments) != null) {
                     if (toolCall2.function.arguments.length > 0) {
                       controller.enqueue({
                         type: "tool-input-delta",
@@ -1008,7 +1102,7 @@ var OpenAIChatLanguageModel = class {
                       });
                       controller.enqueue({
                         type: "tool-call",
-                        toolCallId: (_q = toolCall2.id) != null ? _q : generateId(),
+                        toolCallId: (_r = toolCall2.id) != null ? _r : generateId(),
                         toolName: toolCall2.function.name,
                         input: toolCall2.function.arguments
                       });
@@ -1021,22 +1115,22 @@ var OpenAIChatLanguageModel = class {
                 if (toolCall.hasFinished) {
                   continue;
                 }
-                if (((_r = toolCallDelta.function) == null ? void 0 : _r.arguments) != null) {
-                  toolCall.function.arguments += (_t = (_s = toolCallDelta.function) == null ? void 0 : _s.arguments) != null ? _t : "";
+                if (((_s = toolCallDelta.function) == null ? void 0 : _s.arguments) != null) {
+                  toolCall.function.arguments += (_u = (_t = toolCallDelta.function) == null ? void 0 : _t.arguments) != null ? _u : "";
                 }
                 controller.enqueue({
                   type: "tool-input-delta",
                   id: toolCall.id,
-                  delta: (_u = toolCallDelta.function.arguments) != null ? _u : ""
+                  delta: (_v = toolCallDelta.function.arguments) != null ? _v : ""
                 });
-                if (((_v = toolCall.function) == null ? void 0 : _v.name) != null && ((_w = toolCall.function) == null ? void 0 : _w.arguments) != null && isParsableJson(toolCall.function.arguments)) {
+                if (((_w = toolCall.function) == null ? void 0 : _w.name) != null && ((_x = toolCall.function) == null ? void 0 : _x.arguments) != null && isParsableJson(toolCall.function.arguments)) {
                   controller.enqueue({
                     type: "tool-input-end",
                     id: toolCall.id
                   });
                   controller.enqueue({
                     type: "tool-call",
-                    toolCallId: (_x = toolCall.id) != null ? _x : generateId(),
+                    toolCallId: (_y = toolCall.id) != null ? _y : generateId(),
                     toolName: toolCall.function.name,
                     input: toolCall.function.arguments
                   });
@@ -2304,6 +2398,10 @@ var localShell = createProviderDefinedToolFactoryWithOutputSchema({
 });
 
 // src/responses/convert-to-openai-responses-input.ts
+function getPromptCacheBreakpoint2(providerOptions) {
+  var _a;
+  return (_a = providerOptions == null ? void 0 : providerOptions.openai) == null ? void 0 : _a.promptCacheBreakpoint;
+}
 function isFileId(data, prefixes) {
   if (!prefixes) return false;
   return prefixes.some((prefix) => data.startsWith(prefix));
@@ -2318,16 +2416,36 @@ async function convertToOpenAIResponsesInput({
   var _a, _b, _c, _d, _e, _f;
   let input = [];
   const warnings = [];
-  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);
+            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);
+            input.push({
+              role: "developer",
+              content: promptCacheBreakpoint == null ? content : [
+                {
+                  type: "input_text",
+                  text: content,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ]
+            });
             break;
           }
           case "remove": {
@@ -2353,9 +2471,21 @@ async function convertToOpenAIResponsesInput({
             var _a2, _b2, _c2;
             switch (part.type) {
               case "text": {
-                return { type: "input_text", text: part.text };
+                const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                  part.providerOptions
+                );
+                return {
+                  type: "input_text",
+                  text: part.text,
+                  ...promptCacheBreakpoint != null && {
+                    prompt_cache_breakpoint: promptCacheBreakpoint
+                  }
+                };
               }
               case "file": {
+                const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+                  part.providerOptions
+                );
                 if (part.mediaType.startsWith("image/")) {
                   const mediaType = part.mediaType === "image/*" ? "image/jpeg" : part.mediaType;
                   return {
@@ -2363,13 +2493,19 @@ 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:${mediaType};base64,${convertToBase642(part.data)}`
                     },
-                    detail: (_b2 = (_a2 = part.providerOptions) == null ? void 0 : _a2.openai) == null ? void 0 : _b2.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 === "application/pdf") {
                   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
+                      }
                     };
                   }
                   return {
@@ -2377,6 +2513,9 @@ async function convertToOpenAIResponsesInput({
                     ...typeof part.data === "string" && isFileId(part.data, fileIdPrefixes) ? { file_id: part.data } : {
                       filename: (_c2 = part.filename) != null ? _c2 : `part-${index}.pdf`,
                       file_data: `data:application/pdf;base64,${convertToBase642(part.data)}`
+                    },
+                    ...promptCacheBreakpoint != null && {
+                      prompt_cache_breakpoint: promptCacheBreakpoint
                     }
                   };
                 } else {
@@ -2462,12 +2601,12 @@ async function convertToOpenAIResponsesInput({
               break;
             }
             case "reasoning": {
-              const providerOptions = await parseProviderOptions7({
+              const providerOptions2 = await parseProviderOptions7({
                 provider: "openai",
                 providerOptions: part.providerOptions,
                 schema: openaiResponsesReasoningProviderOptionsSchema
               });
-              const reasoningId = providerOptions == null ? void 0 : providerOptions.itemId;
+              const reasoningId = providerOptions2 == null ? void 0 : providerOptions2.itemId;
               if (reasoningId != null) {
                 const reasoningMessage = reasoningMessages[reasoningId];
                 if (store) {
@@ -2496,14 +2635,14 @@ 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;
                     }
                   }
                 }
@@ -2522,6 +2661,9 @@ async function convertToOpenAIResponsesInput({
       case "tool": {
         for (const part of content) {
           const output = part.output;
+          const promptCacheBreakpoint = getPromptCacheBreakpoint2(
+            part.providerOptions
+          );
           if (hasLocalShellTool && part.toolName === "local_shell" && output.type === "json") {
             const parsedOutput = await validateTypes({
               value: output.value,
@@ -2538,26 +2680,51 @@ async function convertToOpenAIResponsesInput({
           switch (output.type) {
             case "text":
             case "error-text":
-              contentValue = output.value;
+              contentValue = promptCacheBreakpoint == null ? output.value : [
+                {
+                  type: "input_text",
+                  text: output.value,
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ];
               break;
             case "json":
             case "error-json":
-              contentValue = JSON.stringify(output.value);
+              contentValue = promptCacheBreakpoint == null ? JSON.stringify(output.value) : [
+                {
+                  type: "input_text",
+                  text: JSON.stringify(output.value),
+                  prompt_cache_breakpoint: promptCacheBreakpoint
+                }
+              ];
               break;
             case "content":
-              contentValue = output.value.map((item) => {
+              contentValue = output.value.map((item, index) => {
+                const isBreakpoint = promptCacheBreakpoint != null && index === output.value.length - 1;
                 switch (item.type) {
                   case "text": {
-                    return { type: "input_text", text: item.text };
+                    return {
+                      type: "input_text",
+                      text: item.text,
+                      ...isBreakpoint && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
+                    };
                   }
                   case "media": {
                     return item.mediaType.startsWith("image/") ? {
                       type: "input_image",
-                      image_url: `data:${item.mediaType};base64,${item.data}`
+                      image_url: `data:${item.mediaType};base64,${item.data}`,
+                      ...isBreakpoint && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     } : {
                       type: "input_file",
                       filename: "data",
-                      file_data: `data:${item.mediaType};base64,${item.data}`
+                      file_data: `data:${item.mediaType};base64,${item.data}`,
+                      ...isBreakpoint && {
+                        prompt_cache_breakpoint: promptCacheBreakpoint
+                      }
                     };
                   }
                 }
@@ -2648,6 +2815,7 @@ var openaiResponsesChunkSchema = lazyValidator12(
             input_tokens: z15.number(),
             input_tokens_details: z15.object({
               cached_tokens: z15.number().nullish(),
+              cache_write_tokens: z15.number().nullish(),
               orchestration_input_tokens: z15.number().nullish(),
               orchestration_input_cached_tokens: z15.number().nullish()
             }).nullish(),
@@ -2657,6 +2825,9 @@ var openaiResponsesChunkSchema = lazyValidator12(
               orchestration_output_tokens: z15.number().nullish()
             }).nullish()
           }),
+          reasoning: z15.object({
+            context: z15.string().nullish()
+          }).nullish(),
           service_tier: z15.string().nullish()
         })
       }),
@@ -3075,11 +3246,15 @@ var openaiResponsesResponseSchema = lazyValidator12(
         ])
       ).optional(),
       service_tier: z15.string().nullish(),
+      reasoning: z15.object({
+        context: z15.string().nullish()
+      }).nullish(),
       incomplete_details: z15.object({ reason: z15.string() }).nullish(),
       usage: z15.object({
         input_tokens: z15.number(),
         input_tokens_details: z15.object({
           cached_tokens: z15.number().nullish(),
+          cache_write_tokens: z15.number().nullish(),
           orchestration_input_tokens: z15.number().nullish(),
           orchestration_input_cached_tokens: z15.number().nullish()
         }).nullish(),
@@ -3142,7 +3317,11 @@ var openaiResponsesReasoningModelIds = [
   "gpt-5.4-nano",
   "gpt-5.4-nano-2026-03-17",
   "gpt-5.4-pro",
-  "gpt-5.4-pro-2026-03-05"
+  "gpt-5.4-pro-2026-03-05",
+  "gpt-5.6",
+  "gpt-5.6-luna",
+  "gpt-5.6-sol",
+  "gpt-5.6-terra"
 ];
 var openaiResponsesModelIds = [
   "gpt-4.1",
@@ -3216,11 +3395,22 @@ var openaiResponsesProviderOptionsSchema = lazyValidator13(
       parallelToolCalls: z16.boolean().nullish(),
       previousResponseId: z16.string().nullish(),
       promptCacheKey: z16.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: z16.object({
+        mode: z16.enum(["implicit", "explicit"]).optional(),
+        ttl: z16.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'
        */
@@ -3228,14 +3418,21 @@ var openaiResponsesProviderOptionsSchema = lazyValidator13(
       /**
        * 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: z16.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: z16.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: z16.enum(["auto", "current_turn", "all_turns"]).optional(),
       reasoningSummary: z16.string().nullish(),
       safetyIdentifier: z16.string().nullish(),
       serviceTier: z16.enum(["auto", "flex", "priority", "default"]).nullish(),
@@ -3784,18 +3981,25 @@ 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:
-      ...modelCapabilities.isReasoningModel && ((openaiOptions == null ? void 0 : openaiOptions.reasoningEffort) != null || (openaiOptions == null ? void 0 : openaiOptions.reasoningSummary) != null) && {
+      ...modelCapabilities.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
           }
         }
       }
@@ -3834,6 +4038,20 @@ var OpenAIResponsesLanguageModel = class {
           details: "reasoningSummary is not supported for non-reasoning models"
         });
       }
+      if ((openaiOptions == null ? void 0 : openaiOptions.reasoningMode) != null) {
+        warnings.push({
+          type: "unsupported-setting",
+          setting: "reasoningMode",
+          details: "reasoningMode is not supported for non-reasoning models"
+        });
+      }
+      if ((openaiOptions == null ? void 0 : openaiOptions.reasoningContext) != null) {
+        warnings.push({
+          type: "unsupported-setting",
+          setting: "reasoningContext",
+          details: "reasoningContext is not supported for non-reasoning models"
+        });
+      }
     }
     if ((openaiOptions == null ? void 0 : openaiOptions.serviceTier) === "flex" && !modelCapabilities.supportsFlexProcessing) {
       warnings.push({
@@ -3872,7 +4090,7 @@ 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,
@@ -4156,23 +4374,26 @@ var OpenAIResponsesLanguageModel = class {
     if (typeof response.service_tier === "string") {
       providerMetadata[providerKey].serviceTier = response.service_tier;
     }
+    if (((_x = response.reasoning) == null ? void 0 : _x.context) != null) {
+      providerMetadata[providerKey].reasoningContext = response.reasoning.context;
+    }
     const usage = response.usage;
-    const orchestrationUsage = getOrchestrationUsageMetadata(usage);
-    if (orchestrationUsage != null) {
-      providerMetadata[providerKey].usage = orchestrationUsage;
+    const responsesUsage = getResponsesUsageMetadata(usage);
+    if (responsesUsage != null) {
+      providerMetadata[providerKey].usage = responsesUsage;
     }
     return {
       content,
       finishReason: mapOpenAIResponseFinishReason({
-        finishReason: (_x = response.incomplete_details) == null ? void 0 : _x.reason,
+        finishReason: (_y = response.incomplete_details) == null ? void 0 : _y.reason,
         hasFunctionCall
       }),
       usage: {
         inputTokens: usage.input_tokens,
         outputTokens: usage.output_tokens,
         totalTokens: usage.input_tokens + usage.output_tokens,
-        reasoningTokens: (_z = (_y = usage.output_tokens_details) == null ? void 0 : _y.reasoning_tokens) != null ? _z : void 0,
-        cachedInputTokens: (_B = (_A = usage.input_tokens_details) == null ? void 0 : _A.cached_tokens) != null ? _B : void 0
+        reasoningTokens: (_A = (_z = usage.output_tokens_details) == null ? void 0 : _z.reasoning_tokens) != null ? _A : void 0,
+        cachedInputTokens: (_C = (_B = usage.input_tokens_details) == null ? void 0 : _B.cached_tokens) != null ? _C : void 0
       },
       request: { body },
       response: {
@@ -4226,7 +4447,8 @@ var OpenAIResponsesLanguageModel = class {
     let hasFunctionCall = false;
     const activeReasoning = {};
     let serviceTier;
-    let orchestrationUsage;
+    let reasoningContext;
+    let responsesUsage;
     return {
       stream: response.pipeThrough(
         new TransformStream({
@@ -4234,7 +4456,7 @@ 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;
+            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;
             if (options.includeRawChunks) {
               controller.enqueue({ type: "raw", rawValue: chunk.rawValue });
             }
@@ -4627,16 +4849,17 @@ var OpenAIResponsesLanguageModel = class {
               if (typeof value.response.service_tier === "string") {
                 serviceTier = value.response.service_tier;
               }
-              orchestrationUsage = getOrchestrationUsageMetadata(
-                value.response.usage
-              );
+              if (((_p = value.response.reasoning) == null ? void 0 : _p.context) != null) {
+                reasoningContext = value.response.reasoning.context;
+              }
+              responsesUsage = getResponsesUsageMetadata(value.response.usage);
             } else if (isResponseAnnotationAddedChunk(value)) {
               ongoingAnnotations.push(value.annotation);
               if (value.annotation.type === "url_citation") {
                 controller.enqueue({
                   type: "source",
                   sourceType: "url",
-                  id: (_r = (_q = (_p = self.config).generateId) == null ? void 0 : _q.call(_p)) != null ? _r : generateId2(),
+                  id: (_s = (_r = (_q = self.config).generateId) == null ? void 0 : _r.call(_q)) != null ? _s : generateId2(),
                   url: value.annotation.url,
                   title: value.annotation.title
                 });
@@ -4644,10 +4867,10 @@ var OpenAIResponsesLanguageModel = class {
                 controller.enqueue({
                   type: "source",
                   sourceType: "document",
-                  id: (_u = (_t = (_s = self.config).generateId) == null ? void 0 : _t.call(_s)) != null ? _u : generateId2(),
+                  id: (_v = (_u = (_t = self.config).generateId) == null ? void 0 : _u.call(_t)) != null ? _v : generateId2(),
                   mediaType: "text/plain",
-                  title: (_w = (_v = value.annotation.quote) != null ? _v : value.annotation.filename) != null ? _w : "Document",
-                  filename: (_x = value.annotation.filename) != null ? _x : value.annotation.file_id,
+                  title: (_x = (_w = value.annotation.quote) != null ? _w : value.annotation.filename) != null ? _x : "Document",
+                  filename: (_y = value.annotation.filename) != null ? _y : value.annotation.file_id,
                   ...value.annotation.file_id ? {
                     providerMetadata: {
                       [providerKey]: {
@@ -4673,8 +4896,11 @@ var OpenAIResponsesLanguageModel = class {
             if (serviceTier !== void 0) {
               providerMetadata[providerKey].serviceTier = serviceTier;
             }
-            if (orchestrationUsage != null) {
-              providerMetadata[providerKey].usage = orchestrationUsage;
+            if (reasoningContext !== void 0) {
+              providerMetadata[providerKey].reasoningContext = reasoningContext;
+            }
+            if (responsesUsage != null) {
+              providerMetadata[providerKey].usage = responsesUsage;
             }
             controller.enqueue({
               type: "finish",
@@ -4744,15 +4970,17 @@ function mapWebSearchOutput(action) {
       };
   }
 }
-function getOrchestrationUsageMetadata(usage) {
-  var _a, _b, _c;
-  const orchestrationInputTokens = (_a = usage == null ? void 0 : usage.input_tokens_details) == null ? void 0 : _a.orchestration_input_tokens;
-  const orchestrationInputCachedTokens = (_b = usage == null ? void 0 : usage.input_tokens_details) == null ? void 0 : _b.orchestration_input_cached_tokens;
-  const orchestrationOutputTokens = (_c = usage == null ? void 0 : usage.output_tokens_details) == null ? void 0 : _c.orchestration_output_tokens;
-  if (orchestrationInputTokens == null && orchestrationInputCachedTokens == null && orchestrationOutputTokens == null) {
+function getResponsesUsageMetadata(usage) {
+  var _a, _b, _c, _d;
+  const cacheWriteTokens = (_a = usage == null ? void 0 : usage.input_tokens_details) == null ? void 0 : _a.cache_write_tokens;
+  const orchestrationInputTokens = (_b = usage == null ? void 0 : usage.input_tokens_details) == null ? void 0 : _b.orchestration_input_tokens;
+  const orchestrationInputCachedTokens = (_c = usage == null ? void 0 : usage.input_tokens_details) == null ? void 0 : _c.orchestration_input_cached_tokens;
+  const orchestrationOutputTokens = (_d = usage == null ? void 0 : usage.output_tokens_details) == null ? void 0 : _d.orchestration_output_tokens;
+  if (cacheWriteTokens == null && orchestrationInputTokens == null && orchestrationInputCachedTokens == null && orchestrationOutputTokens == null) {
     return void 0;
   }
   return {
+    ...cacheWriteTokens != null && { cacheWriteTokens },
     ...orchestrationInputTokens != null && { orchestrationInputTokens },
     ...orchestrationInputCachedTokens != null && {
       orchestrationInputCachedTokens