@@ -0,0 +1,84 @@
+// npx vitest run src/__tests__/cache-write-ttl-pricing.test.ts
+
+import { describe, it, expect } from "vitest"
+
+import { anthropicModels } from "../providers/anthropic.js"
+import { bedrockModels } from "../providers/bedrock.js"
+import { openAiNativeModels } from "../providers/openai.js"
+import {
+	CACHE_WRITE_MULTIPLIER_12H,
+	CACHE_WRITE_MULTIPLIER_1H,
+	deriveCacheWritesPrice12h,
+	deriveCacheWritesPrice1h,
+	type ModelInfo,
+} from "../model.js"
+
+/**
+ * Catalog invariant for the TTL accounting program (ADR-0006 phase 1).
+ *
+ * A model that advertises prompt caching on a provider which can report a
+ * long-lived cache write must carry the matching per-TTL write rate. Without it
+ * the cost layer can only fall back to the aggregate write rate, which
+ * under-reports every long-lived write (37.5% for Anthropic 1h, 50% for OpenAI 12h).
+ */
+describe("cache-write TTL pricing catalog invariant", () => {
+	const anthropicFamilyEntries = Object.entries(anthropicModels as Record<string, ModelInfo>)
+
+	const bedrockAnthropicEntries = Object.entries(bedrockModels as Record<string, ModelInfo>).filter(([id]) =>
+		id.startsWith("anthropic."),
+	)
+
+	const openAiLongLivedEntries = Object.entries(openAiNativeModels as Record<string, ModelInfo>).filter(
+		([id]) => id.startsWith("gpt-5.6") || id.startsWith("gpt-6"),
+	)
+
+	it("every Anthropic catalog entry with prompt caching defines cacheWritesPrice1h", () => {
+		const missing = anthropicFamilyEntries
+			.filter(([, info]) => info.supportsPromptCache && info.cacheWritesPrice1h === undefined)
+			.map(([id]) => id)
+
+		expect(missing).toEqual([])
+	})
+
+	it("every Bedrock Anthropic-family entry with prompt caching defines cacheWritesPrice1h", () => {
+		const missing = bedrockAnthropicEntries
+			.filter(([, info]) => info.supportsPromptCache && info.cacheWritesPrice1h === undefined)
+			.map(([id]) => id)
+
+		expect(missing).toEqual([])
+	})
+
+	it("every OpenAI 5.6/6-class entry defines cacheWritesPrice12h", () => {
+		expect(openAiLongLivedEntries.length).toBeGreaterThan(0)
+
+		const missing = openAiLongLivedEntries
+			.filter(([, info]) => info.supportsPromptCache && info.cacheWritesPrice12h === undefined)
+			.map(([id]) => id)
+
+		expect(missing).toEqual([])
+	})
+
+	it("Anthropic 1h rates match the documented 2x input-price relationship", () => {
+		for (const [id, info] of anthropicFamilyEntries) {
+			if (info.cacheWritesPrice1h === undefined || info.inputPrice === undefined) continue
+			expect(info.cacheWritesPrice1h, id).toBeCloseTo(info.inputPrice * CACHE_WRITE_MULTIPLIER_1H, 9)
+		}
+	})
+
+	it("OpenAI 12h rates match the documented 2.5x input-price relationship", () => {
+		for (const [id, info] of openAiLongLivedEntries) {
+			if (info.cacheWritesPrice12h === undefined || info.inputPrice === undefined) continue
+			expect(info.cacheWritesPrice12h, id).toBeCloseTo(info.inputPrice * CACHE_WRITE_MULTIPLIER_12H, 9)
+		}
+	})
+
+	it("deriveCacheWritesPrice1h scales inputPrice by 2x and stays undefined without one", () => {
+		expect(deriveCacheWritesPrice1h(3.0)).toBe(6.0)
+		expect(deriveCacheWritesPrice1h(undefined)).toBeUndefined()
+	})
+
+	it("deriveCacheWritesPrice12h scales inputPrice by 2.5x and stays undefined without one", () => {
+		expect(deriveCacheWritesPrice12h(4.0)).toBe(10.0)
+		expect(deriveCacheWritesPrice12h(undefined)).toBeUndefined()
+	})
+})
@@ -122,6 +122,7 @@ describe("OpenAI native models", () => {
 			inputPrice: 2,
 			outputPrice: 10,
 			cacheWritesPrice: 2.5,
+			cacheWritesPrice12h: 5,
 			cacheReadsPrice: 0.2,
 			longContextPricing: {
 				thresholdTokens: 272_000,
@@ -166,6 +167,7 @@ describe("OpenAI native models", () => {
 			inputPrice: 0.1,
 			outputPrice: 0.5,
 			cacheWritesPrice: 0.125,
+			cacheWritesPrice12h: 0.25,
 			cacheReadsPrice: 0.01,
 			longContextPricing: {
 				thresholdTokens: 272_000,
 
@@ -0,0 +1,102 @@
+// npx vitest run src/__tests__/prompt-cache-ttl.test.ts
+
+import { describe, it, expect } from "vitest"
+
+import { anthropicModels, anthropicProviderDefinition } from "../providers/anthropic.js"
+import { vertexModels } from "../providers/vertex.js"
+import { openRouterDefaultModelInfo } from "../providers/openrouter.js"
+import {
+	vercelAiGatewayDefaultModelId,
+	VERCEL_AI_GATEWAY_PROMPT_CACHING_MODELS,
+} from "../providers/vercel-ai-gateway.js"
+import {
+	DEFAULT_PROMPT_CACHE_TTL,
+	modelInfoSchema,
+	promptCacheTtlSchema,
+	promptCacheTtlValues,
+	type ModelInfo,
+} from "../model.js"
+import { baseProviderSettingsShape } from "../providers/common.js"
+import { providerIdentifiers } from "../provider-identifiers.js"
+import { globalSettingsSchema } from "../global-settings.js"
+
+/**
+ * Catalog invariant for the request-side TTL knob (ADR-0006 phase 2, PR-C5b).
+ *
+ * Emission is gated by `ModelInfo.promptCacheTtlSupport`, so a model that can
+ * report a 1h cache write but carries no capability array would silently never
+ * emit the TTL the user asked for. These tests pin both directions: capable
+ * catalogs advertise it, and catalogs that cannot know stay absent (fail closed).
+ */
+describe("prompt cache TTL schema", () => {
+	it("accepts exactly the 5m and 1h classes", () => {
+		expect(promptCacheTtlValues).toEqual(["5m", "1h"])
+		expect(promptCacheTtlSchema.parse("5m")).toBe("5m")
+		expect(promptCacheTtlSchema.parse("1h")).toBe("1h")
+		expect(promptCacheTtlSchema.safeParse("30m").success).toBe(false)
+		expect(promptCacheTtlSchema.safeParse("12h").success).toBe(false)
+	})
+
+	it("defaults to undefined so an unset knob emits no ttl key", () => {
+		expect(DEFAULT_PROMPT_CACHE_TTL).toBeUndefined()
+	})
+
+	it("is optional on ModelInfo", () => {
+		const parsed = modelInfoSchema.parse({ contextWindow: 200_000, supportsPromptCache: true })
+		expect(parsed.promptCacheTtlSupport).toBeUndefined()
+
+		const withCapability = modelInfoSchema.parse({
+			contextWindow: 200_000,
+			supportsPromptCache: true,
+			promptCacheTtlSupport: ["5m", "1h"],
+		})
+		expect(withCapability.promptCacheTtlSupport).toEqual(["5m", "1h"])
+	})
+})
+
+describe("prompt cache TTL knob location", () => {
+	it("lives in provider settings so handlers can see it", () => {
+		// ApiHandlerOptions = Omit<ProviderSettings, "apiProvider">, so only a
+		// provider-settings key reaches a handler (ADR-0006 decision item 2).
+		expect(baseProviderSettingsShape.promptCacheTtl).toBeDefined()
+
+		const parsed = anthropicProviderDefinition.schema.safeParse({
+			apiProvider: providerIdentifiers.anthropic,
+			promptCacheTtl: "1h",
+		})
+		expect(parsed.success).toBe(true)
+		expect(parsed.success && parsed.data.promptCacheTtl).toBe("1h")
+	})
+
+	it("is not a global setting", () => {
+		const shape = globalSettingsSchema.shape as Record<string, unknown>
+		expect(shape.promptCacheTtl).toBeUndefined()
+	})
+})
+
+describe("prompt cache TTL capability catalog", () => {
+	it("every Anthropic entry with prompt caching advertises both TTL classes", () => {
+		const missing = Object.entries(anthropicModels as Record<string, ModelInfo>)
+			.filter(([, info]) => info.supportsPromptCache && !info.promptCacheTtlSupport?.includes("1h"))
+			.map(([id]) => id)
+
+		expect(missing).toEqual([])
+	})
+
+	it("leaves the Vertex catalog without capability so it fails closed", () => {
+		// PR-C5b scopes the catalog capability to the Anthropic entries; the Vertex
+		// catalog is populated separately, so its models stay 5m-only for now.
+		const withCapability = Object.entries(vertexModels as Record<string, ModelInfo>)
+			.filter(([, info]) => info.promptCacheTtlSupport !== undefined)
+			.map(([id]) => id)
+
+		expect(withCapability).toEqual([])
+	})
+
+	it("leaves dynamic-catalog defaults without capability so they fail closed", () => {
+		// OpenRouter and the gateway catalogs are fetched at runtime and carry no TTL
+		// capability data, which ADR-0006 maps to "5m only" (invariant I2).
+		expect(openRouterDefaultModelInfo.promptCacheTtlSupport).toBeUndefined()
+		expect(VERCEL_AI_GATEWAY_PROMPT_CACHING_MODELS.has(vercelAiGatewayDefaultModelId)).toBe(true)
+	})
+})
@@ -95,6 +95,20 @@ export type ModelParameter = z.infer<typeof modelParametersSchema>
 export const isModelParameter = (value: string): value is ModelParameter =>
 	modelParameters.includes(value as ModelParameter)
 
+/**
+ * Prompt cache TTL classes a request-side cache marker can carry. `"5m"` is the
+ * vendor default and is never emitted explicitly, so only `"1h"` changes the wire
+ * shape and an unset knob stays byte-identical to a pre-knob build (ADR-0006 I3).
+ */
+export const promptCacheTtlValues = ["5m", "1h"] as const
+
+export const promptCacheTtlSchema = z.enum(promptCacheTtlValues)
+
+export type PromptCacheTtl = z.infer<typeof promptCacheTtlSchema>
+
+/** Deliberately `undefined`, not `"5m"`: emitting `ttl: "5m"` would change every request's bytes. */
+export const DEFAULT_PROMPT_CACHE_TTL: PromptCacheTtl | undefined = undefined
+
 /**
  * ModelInfo
  */
@@ -112,6 +126,8 @@ export const modelInfoSchema = z.object({
 	// 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(),
+	// Cache TTLs this model accepts. Absent means "5m only" (ADR-0006 I2, fail closed).
+	promptCacheTtlSupport: z.array(promptCacheTtlSchema).optional(),
 	// Capability flag to indicate whether the model supports an output verbosity parameter
 	supportsVerbosity: z.boolean().optional(),
 	// Capability flag to indicate whether the model exposes a user-configurable max output
@@ -136,6 +152,13 @@ export const modelInfoSchema = z.object({
 	inputPrice: z.number().optional(),
 	outputPrice: z.number().optional(),
 	cacheWritesPrice: z.number().optional(),
+	// Per-TTL cache-write rates. Vendors bill a longer cache lifetime at a higher
+	// write rate than the default 5m/30m class, so a single `cacheWritesPrice`
+	// under-reports whenever the response reports a long-lived write. Omitted
+	// means "no long-lived writes are expected on this model"; the cost layer
+	// derives a fail-closed rate from `inputPrice` if one still arrives.
+	cacheWritesPrice1h: z.number().optional(),
+	cacheWritesPrice12h: z.number().optional(),
 	cacheReadsPrice: z.number().optional(),
 	longContextPricing: z
 		.object({
@@ -180,6 +203,11 @@ export const modelInfoSchema = z.object({
 				inputPrice: z.number().optional(),
 				outputPrice: z.number().optional(),
 				cacheWritesPrice: z.number().optional(),
+				// Per-TTL write rates for this tier. A tier that overrides inputPrice
+				// must also override these, otherwise a long-context or flex/priority
+				// write bills at the base tier's TTL rate.
+				cacheWritesPrice1h: z.number().optional(),
+				cacheWritesPrice12h: z.number().optional(),
 				cacheReadsPrice: z.number().optional(),
 			}),
 		)
@@ -188,6 +216,34 @@ export const modelInfoSchema = z.object({
 
 export type ModelInfo = z.infer<typeof modelInfoSchema>
 
+/**
+ * Multipliers vendors apply to the base input price for a cache write, by TTL class.
+ * Anthropic: 5m = 1.25x, 1h = 2x. OpenAI: 30m = 1.25x, 12h = 2.5x.
+ */
+export const CACHE_WRITE_MULTIPLIER_5M = 1.25
+export const CACHE_WRITE_MULTIPLIER_1H = 2.0
+export const CACHE_WRITE_MULTIPLIER_12H = 2.5
+
+/**
+ * Derive the 1h cache-write rate from the base input price.
+ *
+ * Stated once here so every Anthropic-family catalog entry and every runtime
+ * fallback agrees on the 2x relationship. Returns `undefined` when the model
+ * has no input price to scale, which keeps the cost layer fail-closed: it will
+ * fall back to `cacheWritesPrice` rather than invent a rate.
+ */
+export function deriveCacheWritesPrice1h(inputPrice: number | undefined): number | undefined {
+	return inputPrice === undefined ? undefined : inputPrice * CACHE_WRITE_MULTIPLIER_1H
+}
+
+/**
+ * Derive the 12h cache-write rate from the base input price (OpenAI 12h class).
+ * Same fail-closed contract as {@link deriveCacheWritesPrice1h}.
+ */
+export function deriveCacheWritesPrice12h(inputPrice: number | undefined): number | undefined {
+	return inputPrice === undefined ? undefined : inputPrice * CACHE_WRITE_MULTIPLIER_12H
+}
+
 export type ModelRecord = Record<string, ModelInfo>
 
 export type RouterModels = Record<DynamicProvider | LocalProvider, ModelRecord>