diff --git a/packages/ai/src/generate-text/generate-text.test.ts b/packages/ai/src/generate-text/generate-text.test.ts
index 042631852e0d..7c284ee4b3e2 100644
--- a/packages/ai/src/generate-text/generate-text.test.ts
+++ b/packages/ai/src/generate-text/generate-text.test.ts
@@ -8835,4 +8835,100 @@ describe('generateText', () => {
       expect(events).toEqual(['first', 'second']);
     });
   });
+
+  describe('empty text blocks between reasoning blocks', () => {
+    it('should preserve empty text blocks in step 2 messages when reasoning blocks are present', async () => {
+      let responseCount = 0;
+      const doGenerateCalls: Array<Parameters<
+        ConstructorParameters<typeof MockLanguageModelV4>[0] & {
+          doGenerate: (...args: any[]) => any;
+        }
+      >> = [];
+
+      const result = await generateText({
+        model: new MockLanguageModelV4({
+          doGenerate: async options => {
+            doGenerateCalls.push(options as any);
+            switch (responseCount++) {
+              case 0:
+                return {
+                  ...dummyResponseValues,
+                  content: [
+                    {
+                      type: 'reasoning' as const,
+                      text: 'First reasoning block',
+                      signature: 'sig-1',
+                      providerMetadata: {
+                        bedrock: { signature: 'sig-1' },
+                      },
+                    },
+                    { type: 'text' as const, text: '' },
+                    {
+                      type: 'reasoning' as const,
+                      text: 'Second reasoning block',
+                      signature: 'sig-2',
+                      providerMetadata: {
+                        bedrock: { signature: 'sig-2' },
+                      },
+                    },
+                    { type: 'text' as const, text: 'Visible response' },
+                    {
+                      type: 'tool-call' as const,
+                      toolCallType: 'function' as const,
+                      toolCallId: 'call-1',
+                      toolName: 'myTool',
+                      input: '{ "value": "test" }',
+                    },
+                  ],
+                  finishReason: {
+                    unified: 'tool-calls' as const,
+                    raw: undefined,
+                  },
+                };
+              default:
+                return {
+                  ...dummyResponseValues,
+                  content: [{ type: 'text' as const, text: 'Done.' }],
+                };
+            }
+          },
+        }),
+        tools: {
+          myTool: tool({
+            inputSchema: z.object({ value: z.string() }),
+            execute: async ({ value }) => `${value}-result`,
+          }),
+        },
+        prompt: 'test-input',
+        stopWhen: isStepCount(3),
+      });
+
+      // Step 2 should have been called
+      expect(doGenerateCalls.length).toBe(2);
+
+      // Get the assistant message from step 2's prompt
+      const step2Prompt = doGenerateCalls[1].prompt;
+      const assistantMessage = step2Prompt.find(
+        (m: any) => m.role === 'assistant',
+      );
+
+      expect(assistantMessage).toBeDefined();
+
+      // The assistant message content must preserve the empty text block
+      // between reasoning blocks. Without this, providers like Bedrock
+      // reject with "thinking blocks cannot be modified" because the
+      // content block indices shift.
+      const contentTypes = assistantMessage!.content.map(
+        (p: any) => p.type + (p.text === '' ? '(empty)' : ''),
+      );
+
+      expect(contentTypes).toEqual([
+        'reasoning',
+        'text(empty)',
+        'reasoning',
+        'text',
+        'tool-call',
+      ]);
+    });
+  });
 });
diff --git a/packages/ai/src/generate-text/to-response-messages.test.ts b/packages/ai/src/generate-text/to-response-messages.test.ts
index a92edfb5cf5f..cffad136c0be 100644
--- a/packages/ai/src/generate-text/to-response-messages.test.ts
+++ b/packages/ai/src/generate-text/to-response-messages.test.ts
@@ -703,7 +703,7 @@ describe('toResponseMessages', () => {
     `);
   });
 
-  it('should not append text parts if text is empty string', async () => {
+  it('should preserve empty text parts', async () => {
     const result = await toResponseMessages({
       content: [
         {
@@ -729,6 +729,98 @@ describe('toResponseMessages', () => {
       [
         {
           "content": [
+            {
+              "providerOptions": undefined,
+              "text": "",
+              "type": "text",
+            },
+            {
+              "input": {},
+              "providerExecuted": undefined,
+              "providerOptions": undefined,
+              "toolCallId": "123",
+              "toolName": "testTool",
+              "type": "tool-call",
+            },
+          ],
+          "role": "assistant",
+        },
+      ]
+    `);
+  });
+
+  it('should preserve empty text between reasoning blocks for structural integrity', async () => {
+    const result = await toResponseMessages({
+      content: [
+        {
+          type: 'reasoning',
+          text: 'First reasoning block',
+          providerMetadata: {
+            bedrock: { signature: 'sig-1' },
+          },
+        },
+        {
+          type: 'text',
+          text: '',
+        },
+        {
+          type: 'reasoning',
+          text: 'Second reasoning block',
+          providerMetadata: {
+            bedrock: { signature: 'sig-2' },
+          },
+        },
+        {
+          type: 'text',
+          text: 'Final response',
+        },
+        {
+          type: 'tool-call',
+          toolCallId: '123',
+          toolName: 'testTool',
+          input: {},
+        },
+      ],
+      tools: {
+        testTool: tool({
+          description: 'A test tool',
+          inputSchema: z.object({}),
+        }),
+      },
+    });
+
+    expect(result).toMatchInlineSnapshot(`
+      [
+        {
+          "content": [
+            {
+              "providerOptions": {
+                "bedrock": {
+                  "signature": "sig-1",
+                },
+              },
+              "text": "First reasoning block",
+              "type": "reasoning",
+            },
+            {
+              "providerOptions": undefined,
+              "text": "",
+              "type": "text",
+            },
+            {
+              "providerOptions": {
+                "bedrock": {
+                  "signature": "sig-2",
+                },
+              },
+              "text": "Second reasoning block",
+              "type": "reasoning",
+            },
+            {
+              "providerOptions": undefined,
+              "text": "Final response",
+              "type": "text",
+            },
             {
               "input": {},
               "providerExecuted": undefined,
diff --git a/packages/ai/src/generate-text/to-response-messages.ts b/packages/ai/src/generate-text/to-response-messages.ts
index 40f803ebe2c1..9b9185c5bdb2 100644
--- a/packages/ai/src/generate-text/to-response-messages.ts
+++ b/packages/ai/src/generate-text/to-response-messages.ts
@@ -35,11 +35,6 @@ export async function toResponseMessages<TOOLS extends ToolSet>({
       continue;
     }
 
-    // Skip empty text
-    if (part.type === 'text' && part.text.length === 0) {
-      continue;
-    }
-
     switch (part.type) {
       case 'text':
         content.push({
diff --git a/packages/ai/src/prompt/convert-to-language-model-prompt.test.ts b/packages/ai/src/prompt/convert-to-language-model-prompt.test.ts
index d132d49e5333..1f188ce5e35a 100644
--- a/packages/ai/src/prompt/convert-to-language-model-prompt.test.ts
+++ b/packages/ai/src/prompt/convert-to-language-model-prompt.test.ts
@@ -1227,7 +1227,7 @@ describe('convertToLanguageModelMessage', () => {
     });
 
     describe('text parts', () => {
-      it('should ignore empty text parts when there are no provider options', async () => {
+      it('should preserve empty text parts in assistant messages', async () => {
         const result = convertToLanguageModelMessage({
           message: {
             role: 'assistant',
@@ -1250,6 +1250,10 @@ describe('convertToLanguageModelMessage', () => {
         expect(result).toEqual({
           role: 'assistant',
           content: [
+            {
+              type: 'text',
+              text: '',
+            },
             {
               type: 'tool-call',
               input: {},
diff --git a/packages/ai/src/prompt/convert-to-language-model-prompt.ts b/packages/ai/src/prompt/convert-to-language-model-prompt.ts
index 92d5de7cb5eb..53da1f0987c8 100644
--- a/packages/ai/src/prompt/convert-to-language-model-prompt.ts
+++ b/packages/ai/src/prompt/convert-to-language-model-prompt.ts
@@ -223,13 +223,6 @@ export function convertToLanguageModelMessage({
       return {
         role: 'assistant',
         content: message.content
-          .filter(
-            // remove empty text parts (no text, and no provider options):
-            part =>
-              part.type !== 'text' ||
-              part.text !== '' ||
-              part.providerOptions != null,
-          )
           .filter(
             (
               part,