feat(cost-calculator): prompt caching support

Author: ryoppippi

cost-calculator.test.ts

@@ -35,6 +35,31 @@ describe("CostCalculator", () => {
 		expect(cost.outputCost).toEqual(75);
 	});
 
+	test("should calculate cost with cache tokens for opus-4 model", () => {
+		const models = calculator.listModels();
+		const testModel = models.find((m) => m.includes("claude-opus-4-20250514"));
+
+		expect(testModel).toBeDefined();
+
+		const cost = calculator.calculateCost(
+			testModel as string,
+			/* input tokens */ 1_000_000,
+			/* output tokens */ 1_000_000,
+			/* cache creation tokens */ 1_000_000,
+			/* cache read tokens */ 1_000_000,
+		);
+
+		expect(cost.inputTokens).toBe(1_000_000);
+		expect(cost.outputTokens).toBe(1_000_000);
+		expect(cost.cacheCreationTokens).toBe(1_000_000);
+		expect(cost.cacheReadTokens).toBe(1_000_000);
+		expect(cost.inputCost).toEqual(15);
+		expect(cost.outputCost).toEqual(75);
+		expect(cost.cacheCreationCost).toEqual(18.75);
+		expect(cost.cacheReadCost).toEqual(1.5);
+		expect(cost.totalCost).toEqual(15 + 75 + 18.75 + 1.5);
+	});
+
 	test("should calculate cost for a sonnet-4 model", () => {
 		// Using sonnet-4 model as an example
 		// https://www.anthropic.com/pricing
@@ -57,6 +82,33 @@ describe("CostCalculator", () => {
 		expect(cost.outputCost).toEqual(15);
 	});
 
+	test("should calculate cost with cache tokens for sonnet-4 model", () => {
+		const models = calculator.listModels();
+		const testModel = models.find((m) =>
+			m.includes("claude-sonnet-4-20250514"),
+		);
+
+		expect(testModel).toBeDefined();
+
+		const cost = calculator.calculateCost(
+			testModel as string,
+			/* input tokens */ 1_000_000,
+			/* output tokens */ 1_000_000,
+			/* cache creation tokens */ 1_000_000,
+			/* cache read tokens */ 1_000_000,
+		);
+
+		expect(cost.inputTokens).toBe(1_000_000);
+		expect(cost.outputTokens).toBe(1_000_000);
+		expect(cost.cacheCreationTokens).toBe(1_000_000);
+		expect(cost.cacheReadTokens).toBe(1_000_000);
+		expect(cost.inputCost).toEqual(3);
+		expect(cost.outputCost).toEqual(15);
+		expect(cost.cacheCreationCost).toEqual(3.75);
+		expect(cost.cacheReadCost).toEqual(0.3);
+		expect(cost.totalCost).toEqual(3 + 15 + 3.75 + 0.3);
+	});
+
 	test("should throw error for unknown model", () => {
 		expect(() => {
 			calculator.calculateCost("unknown-model", 1000);

cost-calculator.ts

@@ -5,8 +5,12 @@ import { LiteLLMModelPricesSchema } from "./types";
 const CostCalculationSchema = v.object({
 	inputTokens: v.number(),
 	outputTokens: v.number(),
+	cacheCreationTokens: v.number(),
+	cacheReadTokens: v.number(),
 	inputCost: v.number(),
 	outputCost: v.number(),
+	cacheCreationCost: v.number(),
+	cacheReadCost: v.number(),
 	totalCost: v.number(),
 });
 
@@ -35,6 +39,8 @@ export class CostCalculator {
 		modelName: string,
 		inputTokens: number,
 		outputTokens = 0,
+		cacheCreationTokens = 0,
+		cacheReadTokens = 0,
 	): CostCalculation {
 		const modelSpec = this.modelPrices[modelName];
 
@@ -48,13 +54,24 @@ export class CostCalculator {
 
 		const inputCost = inputTokens * modelSpec.input_cost_per_token;
 		const outputCost = outputTokens * modelSpec.output_cost_per_token;
-		const totalCost = inputCost + outputCost;
+		const cacheCreationCost =
+			cacheCreationTokens *
+			(modelSpec.cache_creation_input_token_cost ||
+				modelSpec.input_cost_per_token);
+		const cacheReadCost =
+			cacheReadTokens * (modelSpec.cache_read_input_token_cost || 0);
+		const totalCost =
+			inputCost + outputCost + cacheCreationCost + cacheReadCost;
 
 		return {
 			inputTokens,
 			outputTokens,
+			cacheCreationTokens,
+			cacheReadTokens,
 			inputCost,
 			outputCost,
+			cacheCreationCost,
+			cacheReadCost,
 			totalCost,
 		};
 	}