diff --git a/packages/types/src/model.ts b/packages/types/src/model.ts
index d2f88c0d8ae..df43de936b0 100644
--- a/packages/types/src/model.ts
+++ b/packages/types/src/model.ts
@@ -74,6 +74,10 @@ export const modelInfoSchema = z.object({
 	contextWindow: z.number(),
 	supportsImages: z.boolean().optional(),
 	supportsPromptCache: z.boolean(),
+	// Optional default prompt cache retention policy for providers that support it.
+	// When set to "24h", extended prompt caching will be requested; when omitted
+	// or set to "in_memory", the default in‑memory cache is used.
+	promptCacheRetention: z.enum(["in_memory", "24h"]).optional(),
 	// Capability flag to indicate whether the model supports an output verbosity parameter
 	supportsVerbosity: z.boolean().optional(),
 	supportsReasoningBudget: z.boolean().optional(),
diff --git a/packages/types/src/providers/openai.ts b/packages/types/src/providers/openai.ts
index 77cfaccaae7..76e81ef2dc5 100644
--- a/packages/types/src/providers/openai.ts
+++ b/packages/types/src/providers/openai.ts
@@ -11,6 +11,7 @@ export const openAiNativeModels = {
 		contextWindow: 400000,
 		supportsImages: true,
 		supportsPromptCache: true,
+		promptCacheRetention: "24h",
 		supportsReasoningEffort: ["none", "low", "medium", "high"],
 		reasoningEffort: "medium",
 		inputPrice: 1.25,
@@ -29,6 +30,7 @@ export const openAiNativeModels = {
 		contextWindow: 400000,
 		supportsImages: true,
 		supportsPromptCache: true,
+		promptCacheRetention: "24h",
 		supportsReasoningEffort: ["low", "medium", "high"],
 		reasoningEffort: "medium",
 		inputPrice: 1.25,
@@ -43,6 +45,7 @@ export const openAiNativeModels = {
 		contextWindow: 400000,
 		supportsImages: true,
 		supportsPromptCache: true,
+		promptCacheRetention: "24h",
 		supportsReasoningEffort: ["low", "medium", "high"],
 		reasoningEffort: "medium",
 		inputPrice: 0.25,
diff --git a/src/api/providers/__tests__/openai-native-usage.spec.ts b/src/api/providers/__tests__/openai-native-usage.spec.ts
index 74806b26ab2..48e1c26877b 100644
--- a/src/api/providers/__tests__/openai-native-usage.spec.ts
+++ b/src/api/providers/__tests__/openai-native-usage.spec.ts
@@ -344,6 +344,51 @@ describe("OpenAiNativeHandler - normalizeUsage", () => {
 		})
 	})
 
+	describe("OpenAiNativeHandler - prompt cache retention", () => {
+		let handler: OpenAiNativeHandler
+
+		beforeEach(() => {
+			handler = new OpenAiNativeHandler({
+				openAiNativeApiKey: "test-key",
+			})
+		})
+
+		const buildRequestBodyForModel = (modelId: string) => {
+			// Force the handler to use the requested model ID
+			;(handler as any).options.apiModelId = modelId
+			const model = handler.getModel()
+			// Minimal formatted input/systemPrompt/verbosity/metadata for building the body
+			return (handler as any).buildRequestBody(model, [], "", model.verbosity, undefined, undefined)
+		}
+
+		it("should set prompt_cache_retention=24h for gpt-5.1 models that support prompt caching", () => {
+			const body = buildRequestBodyForModel("gpt-5.1")
+			expect(body.prompt_cache_retention).toBe("24h")
+
+			const codexBody = buildRequestBodyForModel("gpt-5.1-codex")
+			expect(codexBody.prompt_cache_retention).toBe("24h")
+
+			const codexMiniBody = buildRequestBodyForModel("gpt-5.1-codex-mini")
+			expect(codexMiniBody.prompt_cache_retention).toBe("24h")
+		})
+
+		it("should not set prompt_cache_retention for non-gpt-5.1 models even if they support prompt caching", () => {
+			const body = buildRequestBodyForModel("gpt-5")
+			expect(body.prompt_cache_retention).toBeUndefined()
+
+			const fourOBody = buildRequestBodyForModel("gpt-4o")
+			expect(fourOBody.prompt_cache_retention).toBeUndefined()
+		})
+
+		it("should not set prompt_cache_retention when the model does not support prompt caching", () => {
+			const modelId = "codex-mini-latest"
+			expect(openAiNativeModels[modelId as keyof typeof openAiNativeModels].supportsPromptCache).toBe(false)
+
+			const body = buildRequestBodyForModel(modelId)
+			expect(body.prompt_cache_retention).toBeUndefined()
+		})
+	})
+
 	describe("cost calculation", () => {
 		it("should pass total input tokens to calculateApiCostOpenAI", () => {
 			const usage = {
diff --git a/src/api/providers/openai-native.ts b/src/api/providers/openai-native.ts
index 2d51c92a2b3..accae66f271 100644
--- a/src/api/providers/openai-native.ts
+++ b/src/api/providers/openai-native.ts
@@ -52,9 +52,10 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 	constructor(options: ApiHandlerOptions) {
 		super()
 		this.options = options
-		// Default to including reasoning.summary: "auto" for GPT‑5 unless explicitly disabled
-		if (this.options.enableGpt5ReasoningSummary === undefined) {
-			this.options.enableGpt5ReasoningSummary = true
+		// Default to including reasoning.summary: "auto" for models that support Responses API
+		// reasoning summaries unless explicitly disabled.
+		if (this.options.enableResponsesReasoningSummary === undefined) {
+			this.options.enableResponsesReasoningSummary = true
 		}
 		const apiKey = this.options.openAiNativeApiKey ?? "not-provided"
 		this.client = new OpenAI({ baseURL: this.options.openAiNativeBaseUrl, apiKey })
@@ -176,10 +177,10 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		reasoningEffort: ReasoningEffortExtended | undefined,
 		metadata?: ApiHandlerCreateMessageMetadata,
 	): any {
-		// Build a request body
-		// Ensure we explicitly pass max_output_tokens for GPT‑5 based on Roo's reserved model response calculation
+		// Build a request body for the OpenAI Responses API.
+		// Ensure we explicitly pass max_output_tokens based on Roo's reserved model response calculation
 		// so requests do not default to very large limits (e.g., 120k).
-		interface Gpt5RequestBody {
+		interface ResponsesRequestBody {
 			model: string
 			input: Array<{ role: "user" | "assistant"; content: any[] } | { type: string; content: string }>
 			stream: boolean
@@ -191,13 +192,18 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 			instructions?: string
 			service_tier?: ServiceTier
 			include?: string[]
+			/** Prompt cache retention policy: "in_memory" (default) or "24h" for extended caching */
+			prompt_cache_retention?: "in_memory" | "24h"
 		}
 
 		// Validate requested tier against model support; if not supported, omit.
 		const requestedTier = (this.options.openAiNativeServiceTier as ServiceTier | undefined) || undefined
 		const allowedTierNames = new Set(model.info.tiers?.map((t) => t.name).filter(Boolean) || [])
 
-		const body: Gpt5RequestBody = {
+		// Decide whether to enable extended prompt cache retention for this request
+		const promptCacheRetention = this.getPromptCacheRetention(model)
+
+		const body: ResponsesRequestBody = {
 			model: model.id,
 			input: formattedInput,
 			stream: true,
@@ -213,7 +219,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 				? {
 						reasoning: {
 							...(reasoningEffort ? { effort: reasoningEffort } : {}),
-							...(this.options.enableGpt5ReasoningSummary ? { summary: "auto" as const } : {}),
+							...(this.options.enableResponsesReasoningSummary ? { summary: "auto" as const } : {}),
 						},
 					}
 				: {}),
@@ -229,6 +235,9 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 				(requestedTier === "default" || allowedTierNames.has(requestedTier)) && {
 					service_tier: requestedTier,
 				}),
+			// Enable extended prompt cache retention for models that support it.
+			// This uses the OpenAI Responses API `prompt_cache_retention` parameter.
+			...(promptCacheRetention ? { prompt_cache_retention: promptCacheRetention } : {}),
 		}
 
 		// Include text.verbosity only when the model explicitly supports it
@@ -263,7 +272,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 			}
 		} catch (sdkErr: any) {
 			// For errors, fallback to manual SSE via fetch
-			yield* this.makeGpt5ResponsesAPIRequest(requestBody, model, metadata, systemPrompt, messages)
+			yield* this.makeResponsesApiRequest(requestBody, model, metadata, systemPrompt, messages)
 		}
 	}
 
@@ -322,7 +331,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		return formattedMessages
 	}
 
-	private async *makeGpt5ResponsesAPIRequest(
+	private async *makeResponsesApiRequest(
 		requestBody: any,
 		model: OpenAiNativeModel,
 		metadata?: ApiHandlerCreateMessageMetadata,
@@ -347,7 +356,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 			if (!response.ok) {
 				const errorText = await response.text()
 
-				let errorMessage = `GPT-5 API request failed (${response.status})`
+				let errorMessage = `OpenAI Responses API request failed (${response.status})`
 				let errorDetails = ""
 
 				// Try to parse error as JSON for better error messages
@@ -803,7 +812,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 									}
 								}
 
-								// Usage for done/completed is already handled by processGpt5Event in SDK path.
+								// Usage for done/completed is already handled by processEvent in the SDK path.
 								// For SSE path, usage often arrives separately; avoid double-emitting here.
 							}
 							// These are structural or status events, we can just log them at a lower level or ignore.
@@ -977,6 +986,23 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		return selected && selected !== "disable" ? (selected as any) : undefined
 	}
 
+	/**
+	 * Returns the appropriate prompt cache retention policy for the given model, if any.
+	 *
+	 * The policy is driven by ModelInfo.promptCacheRetention so that model-specific details
+	 * live in the shared types layer rather than this provider. When set to "24h" and the
+	 * model supports prompt caching, extended prompt cache retention is requested.
+	 */
+	private getPromptCacheRetention(model: OpenAiNativeModel): "24h" | undefined {
+		if (!model.info.supportsPromptCache) return undefined
+
+		if (model.info.promptCacheRetention === "24h") {
+			return "24h"
+		}
+
+		return undefined
+	}
+
 	/**
 	 * Returns a shallow-cloned ModelInfo with pricing overridden for the given tier, if available.
 	 * If no tier or no overrides exist, the original ModelInfo is returned.
@@ -1083,7 +1109,7 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 			if (reasoningEffort) {
 				requestBody.reasoning = {
 					effort: reasoningEffort,
-					...(this.options.enableGpt5ReasoningSummary ? { summary: "auto" as const } : {}),
+					...(this.options.enableResponsesReasoningSummary ? { summary: "auto" as const } : {}),
 				}
 			}
 
@@ -1102,6 +1128,12 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 				requestBody.text = { verbosity: (verbosity || "medium") as VerbosityLevel }
 			}
 
+			// Enable extended prompt cache retention for eligible models
+			const promptCacheRetention = this.getPromptCacheRetention(model)
+			if (promptCacheRetention) {
+				requestBody.prompt_cache_retention = promptCacheRetention
+			}
+
 			// Make the non-streaming request
 			const response = await (this.client as any).responses.create(requestBody)
 
diff --git a/src/shared/api.ts b/src/shared/api.ts
index 101d7f9b668..4f4c8a4ae9a 100644
--- a/src/shared/api.ts
+++ b/src/shared/api.ts
@@ -13,11 +13,11 @@ import {
 // Extend ProviderSettings (minus apiProvider) with handler-specific toggles.
 export type ApiHandlerOptions = Omit<ProviderSettings, "apiProvider"> & {
 	/**
-	 * When true and using GPT‑5 Responses API, include reasoning.summary: "auto"
-	 * so the API returns reasoning summaries (we already parse and surface them).
-	 * Defaults to true; set to false to disable summaries.
+	 * When true and using OpenAI Responses API models that support reasoning summaries,
+	 * include reasoning.summary: "auto" so the API returns summaries (we already parse
+	 * and surface them). Defaults to true; set to false to disable summaries.
 	 */
-	enableGpt5ReasoningSummary?: boolean
+	enableResponsesReasoningSummary?: boolean
 	/**
 	 * Optional override for Ollama's num_ctx parameter.
 	 * When set, this value will be used in Ollama chat requests.