fix(cost): account cache writes per TTL class instead of one aggregate rate

packages/types/src/__tests__/cache-write-ttl-pricing.test.ts

@@ -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()
+	})
+})

packages/types/src/__tests__/openai-models.test.ts

@@ -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,

packages/types/src/model.ts

@@ -136,6 +136,13 @@ 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 +187,11 @@ 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 +200,34 @@ 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>

packages/types/src/providers/anthropic.ts

@@ -16,6 +16,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens (≤200K context)
 		outputPrice: 15.0, // $15 per million output tokens (≤200K context)
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 		supportsReasoningBudget: true,
 		// Tiered pricing for extended context (requires beta flag 'context-1m-2025-08-07')
@@ -25,6 +26,7 @@ export const anthropicModels = {
 				inputPrice: 6.0, // $6 per million input tokens (>200K context)
 				outputPrice: 22.5, // $22.50 per million output tokens (>200K context)
 				cacheWritesPrice: 7.5, // $7.50 per million tokens (>200K context)
+				cacheWritesPrice1h: 12, // $12.00 per million tokens (2x tier input price, 1h cache writes)
 				cacheReadsPrice: 0.6, // $0.60 per million tokens (>200K context)
 			},
 		],
@@ -37,6 +39,7 @@ export const anthropicModels = {
 		inputPrice: 2.0, // $2 per million input tokens (introductory pricing through Aug 31, 2026)
 		outputPrice: 10.0, // $10 per million output tokens (introductory pricing through Aug 31, 2026)
 		cacheWritesPrice: 2.5, // $2.50 per million tokens (introductory pricing through Aug 31, 2026)
+		cacheWritesPrice1h: 4, // $4.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.2, // $0.20 per million tokens (introductory pricing through Aug 31, 2026)
 		// Sonnet 5 uses the same adaptive-thinking / binary-toggle convention as
 		// Opus 4.7+ and Fable 5 on the direct Anthropic provider path. Manual
@@ -56,6 +59,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens (≤200K context)
 		outputPrice: 15.0, // $15 per million output tokens (≤200K context)
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 		supportsReasoningBudget: true,
 		// Tiered pricing for extended context (requires beta flag 'context-1m-2025-08-07')
@@ -65,6 +69,7 @@ export const anthropicModels = {
 				inputPrice: 6.0, // $6 per million input tokens (>200K context)
 				outputPrice: 22.5, // $22.50 per million output tokens (>200K context)
 				cacheWritesPrice: 7.5, // $7.50 per million tokens (>200K context)
+				cacheWritesPrice1h: 12, // $12.00 per million tokens (2x tier input price, 1h cache writes)
 				cacheReadsPrice: 0.6, // $0.60 per million tokens (>200K context)
 			},
 		],
@@ -77,6 +82,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens (≤200K context)
 		outputPrice: 15.0, // $15 per million output tokens (≤200K context)
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 		supportsReasoningBudget: true,
 		// Tiered pricing for extended context (requires beta flag 'context-1m-2025-08-07')
@@ -86,6 +92,7 @@ export const anthropicModels = {
 				inputPrice: 6.0, // $6 per million input tokens (>200K context)
 				outputPrice: 22.5, // $22.50 per million output tokens (>200K context)
 				cacheWritesPrice: 7.5, // $7.50 per million tokens (>200K context)
+				cacheWritesPrice1h: 12, // $12.00 per million tokens (2x tier input price, 1h cache writes)
 				cacheReadsPrice: 0.6, // $0.60 per million tokens (>200K context)
 			},
 		],
@@ -98,6 +105,7 @@ export const anthropicModels = {
 		inputPrice: 5.0, // $5 per million input tokens (≤200K context)
 		outputPrice: 25.0, // $25 per million output tokens (≤200K context)
 		cacheWritesPrice: 6.25, // $6.25 per million tokens
+		cacheWritesPrice1h: 10, // $10.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.5, // $0.50 per million tokens
 		supportsReasoningBudget: true,
 		// Tiered pricing for extended context (requires beta flag)
@@ -107,6 +115,7 @@ export const anthropicModels = {
 				inputPrice: 10.0, // $10 per million input tokens (>200K context)
 				outputPrice: 37.5, // $37.50 per million output tokens (>200K context)
 				cacheWritesPrice: 12.5, // $12.50 per million tokens (>200K context)
+				cacheWritesPrice1h: 20, // $20.00 per million tokens (2x tier input price, 1h cache writes)
 				cacheReadsPrice: 1.0, // $1.00 per million tokens (>200K context)
 			},
 		],
@@ -119,6 +128,7 @@ export const anthropicModels = {
 		inputPrice: 5.0, // $5 per million input tokens
 		outputPrice: 25.0, // $25 per million output tokens
 		cacheWritesPrice: 6.25, // $6.25 per million tokens
+		cacheWritesPrice1h: 10, // $10.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.5, // $0.50 per million tokens
 		// Keep the hybrid-reasoning capability so Anthropic token-cap handling and
 		// stored max-token overrides behave the same as before.
@@ -136,6 +146,7 @@ export const anthropicModels = {
 		inputPrice: 5.0, // $5 per million input tokens (regular tier)
 		outputPrice: 25.0, // $25 per million output tokens (regular tier)
 		cacheWritesPrice: 6.25, // $6.25 per million tokens
+		cacheWritesPrice1h: 10, // $10.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.5, // $0.50 per million tokens
 		// 4.8 inherits the adaptive-thinking model introduced in 4.7 — no breaking
 		// API changes. supportsReasoningBudget is kept true so the existing token-cap
@@ -154,6 +165,7 @@ export const anthropicModels = {
 		inputPrice: 5.0, // $5 per million input tokens
 		outputPrice: 25.0, // $25 per million output tokens
 		cacheWritesPrice: 6.25, // $6.25 per million tokens
+		cacheWritesPrice1h: 10, // $10.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.5, // $0.50 per million tokens
 		// Opus 5 uses the same adaptive-thinking / binary-toggle convention as
 		// Opus 4.7+ and Sonnet 5 on the direct Anthropic provider path. Manual
@@ -172,6 +184,7 @@ export const anthropicModels = {
 		inputPrice: 4.0, // $4 per million input tokens
 		outputPrice: 20.0, // $20 per million output tokens
 		cacheWritesPrice: 5.0, // $5 per million tokens
+		cacheWritesPrice1h: 8, // $8.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.2, // $0.20 per million tokens
 		// Opus 5.5 always uses adaptive thinking and rejects manual budget_tokens.
 		supportsReasoningBinary: true,
@@ -192,6 +205,7 @@ export const anthropicModels = {
 		inputPrice: 0.1, // $0.10 per million input tokens
 		outputPrice: 0.5, // $0.50 per million output tokens
 		cacheWritesPrice: 0.125, // $0.125 per million tokens (5m cache writes)
+		cacheWritesPrice1h: 0.2, // $0.20 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.01, // $0.01 per million tokens (cache hits / refreshes)
 		// Haiku 5.5 uses the same adaptive-thinking / binary-toggle convention as
 		// Sonnet 5 and Opus 5 on the direct Anthropic provider path. Manual
@@ -211,6 +225,7 @@ export const anthropicModels = {
 		inputPrice: 10.0, // $10 per million input tokens
 		outputPrice: 50.0, // $50 per million output tokens
 		cacheWritesPrice: 12.5, // $12.50 per million tokens (5-minute cache)
+		cacheWritesPrice1h: 20, // $20.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.25, // $0.25 per million tokens (2.5% of base input price)
 		// Fable 5.1 only accepts adaptive thinking (or an omitted thinking field).
 		// Manual budget_tokens and non-default sampling parameters return a 400.
@@ -228,6 +243,7 @@ export const anthropicModels = {
 		inputPrice: 10.0, // $10 per million input tokens
 		outputPrice: 50.0, // $50 per million output tokens
 		cacheWritesPrice: 12.5, // $12.50 per million tokens
+		cacheWritesPrice1h: 20, // $20.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 1.0, // $1.00 per million tokens
 		// Fable 5 uses the same adaptive-thinking / binary-toggle convention as
 		// Opus 4.7+ on the direct Anthropic provider path.
@@ -245,6 +261,7 @@ export const anthropicModels = {
 		inputPrice: 5.0, // $5 per million input tokens
 		outputPrice: 25.0, // $25 per million output tokens
 		cacheWritesPrice: 6.25, // $6.25 per million tokens
+		cacheWritesPrice1h: 10, // $10.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.5, // $0.50 per million tokens
 		supportsReasoningBudget: true,
 	},
@@ -256,6 +273,7 @@ export const anthropicModels = {
 		inputPrice: 15.0, // $15 per million input tokens
 		outputPrice: 75.0, // $75 per million output tokens
 		cacheWritesPrice: 18.75, // $18.75 per million tokens
+		cacheWritesPrice1h: 30, // $30.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 1.5, // $1.50 per million tokens
 		supportsReasoningBudget: true,
 	},
@@ -267,6 +285,7 @@ export const anthropicModels = {
 		inputPrice: 15.0, // $15 per million input tokens
 		outputPrice: 75.0, // $75 per million output tokens
 		cacheWritesPrice: 18.75, // $18.75 per million tokens
+		cacheWritesPrice1h: 30, // $30.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 1.5, // $1.50 per million tokens
 		supportsReasoningBudget: true,
 	},
@@ -278,6 +297,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens
 		outputPrice: 15.0, // $15 per million output tokens
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 		supportsReasoningBudget: true,
 		requiredReasoningBudget: true,
@@ -290,6 +310,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens
 		outputPrice: 15.0, // $15 per million output tokens
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 	},
 	"claude-3-5-sonnet-20241022": {
@@ -300,6 +321,7 @@ export const anthropicModels = {
 		inputPrice: 3.0, // $3 per million input tokens
 		outputPrice: 15.0, // $15 per million output tokens
 		cacheWritesPrice: 3.75, // $3.75 per million tokens
+		cacheWritesPrice1h: 6, // $6.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.3, // $0.30 per million tokens
 	},
 	"claude-3-5-haiku-20241022": {
@@ -310,6 +332,7 @@ export const anthropicModels = {
 		inputPrice: 1.0,
 		outputPrice: 5.0,
 		cacheWritesPrice: 1.25,
+		cacheWritesPrice1h: 2, // $2.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.1,
 	},
 	"claude-3-opus-20240229": {
@@ -320,6 +343,7 @@ export const anthropicModels = {
 		inputPrice: 15.0,
 		outputPrice: 75.0,
 		cacheWritesPrice: 18.75,
+		cacheWritesPrice1h: 30, // $30.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 1.5,
 	},
 	"claude-3-haiku-20240307": {
@@ -330,6 +354,7 @@ export const anthropicModels = {
 		inputPrice: 0.25,
 		outputPrice: 1.25,
 		cacheWritesPrice: 0.3,
+		cacheWritesPrice1h: 0.5, // $0.50 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.03,
 	},
 	"claude-haiku-4-5-20251001": {
@@ -340,6 +365,7 @@ export const anthropicModels = {
 		inputPrice: 1.0,
 		outputPrice: 5.0,
 		cacheWritesPrice: 1.25,
+		cacheWritesPrice1h: 2, // $2.00 per million tokens (2x input price, 1h cache writes)
 		cacheReadsPrice: 0.1,
 		supportsReasoningBudget: true,
 		description: