diff --git a/changelog.md b/changelog.md
index 315be8b..34fe48f 100644
--- a/changelog.md
+++ b/changelog.md
@@ -3,6 +3,13 @@
 ## Pending
 
 
+### Fixes
+
+- map system prompts to top-level instructions for ChatGPT auth requests (openai)
+  - wrap OpenAI models using ChatGPT auth with middleware that extracts system messages into provider instructions
+  - remove system messages from the serialized prompt payload to match the ChatGPT Codex backend expectations
+  - add test coverage to verify authorization headers, workspace routing, and instruction payload mapping
+
 ## 2026-05-14 - 4.0.0
 
 ### Breaking Changes
diff --git a/test/test.openai-auth.node.ts b/test/test.openai-auth.node.ts
index 4757ab8..3b45163 100644
--- a/test/test.openai-auth.node.ts
+++ b/test/test.openai-auth.node.ts
@@ -199,15 +199,21 @@ tap.test('getModel uses ChatGPT Codex backend for OpenAI ChatGPT auth', async ()
   };
 
   try {
-    await model.doGenerate({
-      prompt: [{ role: 'user', content: [{ type: 'text', text: 'hello' }] }],
-      inputFormat: 'prompt',
-    } as any);
+    await smartai.generateText({
+      model,
+      system: 'system prompt',
+      prompt: 'hello',
+    });
 
     expect(capturedRequest?.url).toEqual('https://chatgpt.com/backend-api/codex/responses');
     expect(getHeader(capturedRequest?.init, 'authorization')).toEqual(`Bearer ${tokenData.accessToken}`);
     expect(getHeader(capturedRequest?.init, 'chatgpt-account-id')).toEqual('workspace-1');
     expect(getHeader(capturedRequest?.init, 'originator')).toEqual('smartai');
+    const capturedBody = JSON.parse(String(capturedRequest?.init?.body));
+    expect(capturedBody.instructions).toEqual('system prompt');
+    expect(capturedBody.input).toEqual([
+      { role: 'user', content: [{ type: 'input_text', text: 'hello' }] },
+    ]);
   } finally {
     globalThis.fetch = originalFetch;
   }
diff --git a/ts/smartai.classes.smartai.ts b/ts/smartai.classes.smartai.ts
index a87aeef..964b304 100644
--- a/ts/smartai.classes.smartai.ts
+++ b/ts/smartai.classes.smartai.ts
@@ -2,6 +2,7 @@ import * as plugins from './plugins.js';
 import type { ISmartAiModelSetup, ISmartAiOptions, LanguageModelV3 } from './smartai.interfaces.js';
 import { createOllamaModel } from './smartai.provider.ollama.js';
 import { createAnthropicCachingMiddleware } from './smartai.middleware.anthropic.js';
+import { createOpenAiChatGptInstructionsMiddleware } from './smartai.middleware.openai.js';
 import { createOpenAiChatGptProviderSettings } from './smartai.auth.openai.js';
 
 /**
@@ -28,7 +29,13 @@ export function getModel(options: ISmartAiOptions): LanguageModelV3 {
           ? createOpenAiChatGptProviderSettings(options.openAiChatGptAuth)
           : { apiKey: options.apiKey },
       );
-      return p(options.model) as LanguageModelV3;
+      const base = p(options.model) as LanguageModelV3;
+      return options.openAiChatGptAuth
+        ? plugins.wrapLanguageModel({
+            model: base,
+            middleware: createOpenAiChatGptInstructionsMiddleware(),
+          }) as unknown as LanguageModelV3
+        : base;
     }
     case 'google': {
       const p = plugins.createGoogleGenerativeAI({ apiKey: options.apiKey });
diff --git a/ts/smartai.middleware.openai.ts b/ts/smartai.middleware.openai.ts
new file mode 100644
index 0000000..0af1d59
--- /dev/null
+++ b/ts/smartai.middleware.openai.ts
@@ -0,0 +1,46 @@
+import type { JSONObject, LanguageModelV3CallOptions, LanguageModelV3Middleware } from '@ai-sdk/provider';
+
+const isNonEmptyString = (value: unknown): value is string => typeof value === 'string' && value.trim().length > 0;
+
+const getSystemInstructions = (prompt: LanguageModelV3CallOptions['prompt']): string | undefined => {
+  const instructions = prompt
+    .filter((message) => message.role === 'system')
+    .map((message) => message.content)
+    .filter(isNonEmptyString);
+
+  return instructions.length > 0 ? instructions.join('\n') : undefined;
+};
+
+/**
+ * ChatGPT's Codex backend requires top-level Responses API instructions.
+ * The standard OpenAI provider otherwise serializes system prompts as input items.
+ */
+export function createOpenAiChatGptInstructionsMiddleware(): LanguageModelV3Middleware {
+  return {
+    specificationVersion: 'v3',
+    transformParams: async ({ params }) => {
+      const instructions = getSystemInstructions(params.prompt);
+      if (!instructions) {
+        return params;
+      }
+
+      const providerOptions = params.providerOptions ?? {};
+      const openAiProviderOptions = providerOptions.openai ?? {};
+      if (isNonEmptyString(openAiProviderOptions.instructions)) {
+        return params;
+      }
+
+      return {
+        ...params,
+        prompt: params.prompt.filter((message) => message.role !== 'system'),
+        providerOptions: {
+          ...providerOptions,
+          openai: {
+            ...openAiProviderOptions,
+            instructions,
+          } as JSONObject,
+        },
+      } satisfies LanguageModelV3CallOptions;
+    },
+  };
+}