diff --git a/packages/ai/src/providers/openai-codex.ts b/packages/ai/src/providers/openai-codex.ts
index e2be0d80f..50c66d804 100644
--- a/packages/ai/src/providers/openai-codex.ts
+++ b/packages/ai/src/providers/openai-codex.ts
@@ -24,6 +24,14 @@ function credentialId(credential: Credential | undefined): string | undefined {
 		: undefined;
 }
 
+function withChatGPTCompat(model: Model<"openai-responses">): Model<"openai-responses"> {
+	return {
+		...model,
+		// ChatGPT subscription requests reject the API-key model's explicit-cache parameter.
+		compat: { ...model.compat, supportsExplicitPromptCacheMode: false },
+	};
+}
+
 function parseCatalog(value: unknown): Model<"openai-responses">[] {
 	if (!value || typeof value !== "object" || !("models" in value) || !Array.isArray(value.models)) {
 		throw new Error("Invalid ChatGPT model catalog");
@@ -57,20 +65,22 @@ function parseCatalog(value: unknown): Model<"openai-responses">[] {
 			item.max_output_tokens > 0
 				? item.max_output_tokens
 				: (known?.maxTokens ?? 16_384);
-		models.push({
-			...known,
-			id: item.slug,
-			name: item.display_name,
-			api: "openai-responses",
-			provider: "openai-codex",
-			baseUrl: BASE_URL,
-			reasoning: known?.reasoning ?? false,
-			input: known?.input ?? ["text"],
-			contextWindow,
-			maxTokens: Math.min(maxTokens, contextWindow),
-			// ChatGPT plan usage is not billed at API-key token prices.
-			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-		});
+		models.push(
+			withChatGPTCompat({
+				...known,
+				id: item.slug,
+				name: item.display_name,
+				api: "openai-responses",
+				provider: "openai-codex",
+				baseUrl: BASE_URL,
+				reasoning: known?.reasoning ?? false,
+				input: known?.input ?? ["text"],
+				contextWindow,
+				maxTokens: Math.min(maxTokens, contextWindow),
+				// ChatGPT plan usage is not billed at API-key token prices.
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+			}),
+		);
 	}
 	return models;
 }
@@ -165,12 +175,14 @@ export function openaiCodexProvider(): Provider<"openai-responses"> {
 					update: () => {
 						activate(owner);
 						if (stored) {
-							catalog = stored.models.filter(
-								(model): model is Model<"openai-responses"> =>
-									model.provider === "openai-codex" &&
-									model.api === "openai-responses" &&
-									model.baseUrl === BASE_URL,
-							);
+							catalog = stored.models
+								.filter(
+									(model): model is Model<"openai-responses"> =>
+										model.provider === "openai-codex" &&
+										model.api === "openai-responses" &&
+										model.baseUrl === BASE_URL,
+								)
+								.map(withChatGPTCompat);
 							catalogOwner = owner;
 						}
 					},
diff --git a/packages/ai/test/chatgpt-inference.test.ts b/packages/ai/test/chatgpt-inference.test.ts
index 9650568a0..df9d8bb71 100644
--- a/packages/ai/test/chatgpt-inference.test.ts
+++ b/packages/ai/test/chatgpt-inference.test.ts
@@ -5,8 +5,9 @@ import { Type } from "typebox";
 import { afterEach, expect, it, vi } from "vitest";
 import { InMemoryCredentialStore } from "../src/auth/credential-store.ts";
 import type { OAuthCredential } from "../src/auth/types.ts";
-import { createModels } from "../src/models.ts";
+import { createModels, hasApi } from "../src/models.ts";
 import { InMemoryModelsStore } from "../src/models-store.ts";
+import { openaiProvider } from "../src/providers/openai.ts";
 import { openaiCodexProvider } from "../src/providers/openai-codex.ts";
 import type { Context } from "../src/types.ts";
 
@@ -167,6 +168,81 @@ it("streams text and usage over public Responses with subscription billing", asy
 	});
 });
 
+it.each(["fresh", "cached offline", "cached after refresh failure"])(
+	"omits unsupported prompt-cache options for summaries with a %s catalog",
+	async (catalogState) => {
+		const bodies: Record<string, unknown>[] = [];
+		let failCatalogRefresh = false;
+		vi.stubGlobal(
+			"fetch",
+			vi.fn(async (input: string | URL | Request, init?: RequestInit) => {
+				if (String(input).endsWith("/models")) {
+					if (failCatalogRefresh) return new Response(null, { status: 503 });
+					return Response.json(catalog);
+				}
+				const body = JSON.parse(String(init?.body)) as Record<string, unknown>;
+				bodies.push(body);
+				if ("prompt_cache_options" in body) {
+					return Response.json(
+						{
+							error: {
+								message: "prompt_cache_options is not supported on this model",
+								type: "invalid_request_error",
+								param: "prompt_cache_options",
+								code: "invalid_parameter",
+							},
+						},
+						{ status: 400 },
+					);
+				}
+				return sse([...textEvents(), completed()]);
+			}),
+		);
+		const initial = await runtime();
+		let models = initial.models;
+		if (catalogState !== "fresh") {
+			const stored = await initial.store.read("openai-codex");
+			if (!stored) throw new Error("Catalog not persisted");
+			// Simulate a catalog persisted before the hotfix with inherited API-key capabilities.
+			await initial.store.write("openai-codex", {
+				...stored,
+				models: stored.models.map((model) => ({
+					...model,
+					compat: { ...model.compat, supportsExplicitPromptCacheMode: true },
+				})),
+			});
+			models = createModels({ credentials: initial.credentials, modelsStore: initial.store });
+			models.setProvider(openaiCodexProvider());
+			failCatalogRefresh = catalogState === "cached after refresh failure";
+			const refresh = await models.refresh({ allowNetwork: failCatalogRefresh });
+			expect(refresh.errors.size).toBe(failCatalogRefresh ? 1 : 0);
+		}
+		const model = models.getModel("openai-codex", "gpt-6.1-sol");
+		if (!model || !hasApi(model, "openai-responses")) throw new Error("Responses model not discovered");
+		const options = { cacheRetention: "none" as const, sessionId: "summary-session" };
+		for (const result of [
+			await models.completeSimple(model, context, options),
+			await models.complete(model, context, options),
+		]) {
+			expect(result.stopReason).toBe("stop");
+			expect(result.content).toEqual([expect.objectContaining({ type: "text", text: "Hello" })]);
+		}
+		expect(model.compat?.supportsExplicitPromptCacheMode).toBe(false);
+		expect(model.compat?.supportsStrictMode).toBe(true);
+		expect(model.compat?.supportsOpenAIGrammarTools).toBe(true);
+		const apiKeyModel = openaiProvider()
+			.getModels()
+			.find((entry) => entry.id === model.id);
+		expect(apiKeyModel?.compat?.supportsExplicitPromptCacheMode).toBe(true);
+		expect(bodies).toHaveLength(2);
+		for (const body of bodies) {
+			expect(body).not.toHaveProperty("prompt_cache_options");
+			expect(body).not.toHaveProperty("prompt_cache_key");
+			expect(body).not.toHaveProperty("prompt_cache_retention");
+		}
+	},
+);
+
 it("streams tool arguments and replays tool results and encrypted reasoning", async () => {
 	const bodies: Record<string, unknown>[] = [];
 	vi.stubGlobal(