@clinebot/llms 0.0.10 → 0.0.11

package/dist/providers/handlers/ai-sdk-community.d.ts

@@ -28,7 +28,7 @@ type AiSdkUsageMetrics = {
 export type EmitAiSdkStreamOptions = {
     responseId: string;
     errorMessage: string;
-    calculateCost: (inputTokens: number, outputTokens: number, cacheReadTokens: number) => number | undefined;
+    calculateCost: (inputTokens: number, outputTokens: number, cacheReadTokens: number, cacheWriteTokens?: number) => number | undefined;
     reasoningTypes?: string[];
     enableToolCalls?: boolean;
     toolCallArgsOrder?: Array<"args" | "input">;

package/dist/providers/handlers/base.d.ts

@@ -3,7 +3,7 @@
  *
  * Abstract base class that provides common functionality for all handlers.
  */
-import type { ApiHandler, ApiStream, ApiStreamUsageChunk, HandlerModelInfo, ProviderConfig } from "../types";
+import type { ApiHandler, ApiStream, ApiStreamUsageChunk, HandlerModelInfo, ModelInfo, ProviderConfig } from "../types";
 import type { Message, ToolDefinition } from "../types/messages";
 import type { ApiStreamChunk } from "../types/stream";
 export declare const DEFAULT_REQUEST_HEADERS: Record<string, string>;
@@ -23,7 +23,8 @@ export declare abstract class BaseHandler implements ApiHandler {
     abort(): void;
     setAbortSignal(signal: AbortSignal | undefined): void;
     private logAbort;
-    protected calculateCost(inputTokens: number, outputTokens: number, cacheReadTokens?: number): number | undefined;
+    protected supportsPromptCache(modelInfo?: ModelInfo): boolean;
+    protected calculateCost(inputTokens: number, outputTokens: number, cacheReadTokens?: number, cacheWriteTokens?: number): number | undefined;
     protected createResponseId(): string;
     protected withResponseId<T extends ApiStreamChunk>(chunk: T, responseId: string): T;
     protected withResponseIdForAll(chunks: Iterable<ApiStreamChunk>, responseId: string): Generator<ApiStreamChunk>;

package/dist/providers/types/stream.d.ts

@@ -47,7 +47,7 @@ export interface ApiStreamReasoningChunk {
  */
 export interface ApiStreamUsageChunk {
     type: "usage";
-    /** Number of input tokens (excluding cached) */
+    /** Total number of input tokens reported by the provider */
     inputTokens: number;
     /** Number of output tokens */
     outputTokens: number;

package/package.json

@@ -1,11 +1,11 @@
 {
 	"name": "@clinebot/llms",
-	"version": "0.0.10",
+	"version": "0.0.11",
 	"description": "Config-driven SDK for selecting, extending, and instantiating LLM providers and models",
 	"main": "./dist/index.js",
 	"module": "./dist/index.js",
 	"dependencies": {
-		"@clinebot/shared": "0.0.10",
+		"@clinebot/shared": "0.0.11",
 		"@ai-sdk/amazon-bedrock": "^4.0.67",
 		"@ai-sdk/google-vertex": "^4.0.74",
 		"@ai-sdk/mistral": "^3.0.24",

package/src/providers/handlers/ai-sdk-community.ts

@@ -41,6 +41,7 @@ export type EmitAiSdkStreamOptions = {
 		inputTokens: number,
 		outputTokens: number,
 		cacheReadTokens: number,
+		cacheWriteTokens?: number,
 	) => number | undefined;
 	reasoningTypes?: string[];
 	enableToolCalls?: boolean;
@@ -168,10 +169,7 @@ export async function* emitAiSdkStream(
 
 				yield {
 					type: "usage",
-					inputTokens: Math.max(
-						0,
-						usageMetrics.inputTokens - usageMetrics.cacheReadTokens,
-					),
+					inputTokens: usageMetrics.inputTokens,
 					outputTokens: usageMetrics.outputTokens,
 					thoughtsTokenCount: usageMetrics.thoughtsTokenCount,
 					cacheReadTokens: usageMetrics.cacheReadTokens,
@@ -180,6 +178,7 @@ export async function* emitAiSdkStream(
 						usageMetrics.inputTokens,
 						usageMetrics.outputTokens,
 						usageMetrics.cacheReadTokens,
+						usageMetrics.cacheWriteTokens,
 					),
 					id: responseId,
 				};
@@ -205,10 +204,7 @@ export async function* emitAiSdkStream(
 		const usageMetrics = resolveUsageMetrics(usage);
 		yield {
 			type: "usage",
-			inputTokens: Math.max(
-				0,
-				usageMetrics.inputTokens - usageMetrics.cacheReadTokens,
-			),
+			inputTokens: usageMetrics.inputTokens,
 			outputTokens: usageMetrics.outputTokens,
 			thoughtsTokenCount: usageMetrics.thoughtsTokenCount,
 			cacheReadTokens: usageMetrics.cacheReadTokens,
@@ -217,6 +213,7 @@ export async function* emitAiSdkStream(
 				usageMetrics.inputTokens,
 				usageMetrics.outputTokens,
 				usageMetrics.cacheReadTokens,
+				usageMetrics.cacheWriteTokens,
 			),
 			id: responseId,
 		};

package/src/providers/handlers/ai-sdk-provider-base.ts

@@ -185,8 +185,18 @@ export abstract class AiSdkProviderHandler extends BaseHandler {
 		yield* emitAiSdkStream(stream, {
 			responseId,
 			errorMessage: this.getStreamErrorMessage(),
-			calculateCost: (inputTokens, outputTokens, cacheReadTokens) =>
-				this.calculateCost(inputTokens, outputTokens, cacheReadTokens),
+			calculateCost: (
+				inputTokens,
+				outputTokens,
+				cacheReadTokens,
+				cacheWriteTokens,
+			) =>
+				this.calculateCost(
+					inputTokens,
+					outputTokens,
+					cacheReadTokens,
+					cacheWriteTokens,
+				),
 			...this.getEmitStreamOptions(),
 		});
 	}

package/src/providers/handlers/anthropic-base.test.ts

@@ -0,0 +1,30 @@
+import { describe, expect, it } from "vitest";
+import { AnthropicHandler } from "./anthropic-base";
+
+describe("AnthropicHandler prompt cache detection", () => {
+	it("enables prompt caching when model pricing includes cache pricing", () => {
+		const handler = new AnthropicHandler({
+			providerId: "anthropic",
+			modelId: "claude-sonnet-4-6",
+			apiKey: "test-key",
+			modelInfo: {
+				id: "claude-sonnet-4-6",
+				pricing: {
+					input: 3,
+					output: 15,
+					cacheRead: 0.3,
+					cacheWrite: 3.75,
+				},
+			},
+		});
+
+		const messages = handler.getMessages("system", [
+			{ role: "user", content: "Tell me about this repo" },
+		]);
+		const userTextBlock = messages[0]?.content?.[0] as
+			| { cache_control?: { type: string } }
+			| undefined;
+
+		expect(userTextBlock?.cache_control).toEqual({ type: "ephemeral" });
+	});
+});

package/src/providers/handlers/anthropic-base.ts

@@ -17,7 +17,6 @@ import {
 import {
 	type ApiStream,
 	type HandlerModelInfo,
-	hasModelCapability,
 	type ProviderConfig,
 	supportsModelThinking,
 } from "../types";
@@ -76,10 +75,7 @@ export class AnthropicHandler extends BaseHandler {
 		_systemPrompt: string,
 		messages: Message[],
 	): Anthropic.MessageParam[] {
-		const supportsPromptCache = hasModelCapability(
-			this.getModel().info,
-			"prompt-cache",
-		);
+		const supportsPromptCache = this.supportsPromptCache(this.getModel().info);
 		return convertToAnthropicMessages(
 			messages,
 			supportsPromptCache,
@@ -113,7 +109,7 @@ export class AnthropicHandler extends BaseHandler {
 		const budgetTokens =
 			thinkingSupported && requestedBudget > 0 ? requestedBudget : 0;
 		const nativeToolsOn = tools && tools.length > 0;
-		const supportsPromptCache = hasModelCapability(model.info, "prompt-cache");
+		const supportsPromptCache = this.supportsPromptCache(model.info);
 		const reasoningOn = thinkingSupported && budgetTokens > 0;
 		const debugThinking = isThinkingDebugEnabled();
 		const debugChunkCounts: Record<string, number> = {};
@@ -139,30 +135,34 @@ export class AnthropicHandler extends BaseHandler {
 		const requestOptions = { signal: abortSignal };
 
 		// Create the request
+		// Use top-level automatic caching so the entire prefix (system +
+		// messages) is cached and the breakpoint advances each turn.
+		const createParams: Record<string, unknown> &
+			Anthropic.MessageCreateParamsStreaming = {
+			model: model.id,
+			thinking: reasoningOn
+				? { type: "enabled", budget_tokens: budgetTokens }
+				: undefined,
+			max_tokens:
+				model.info.maxTokens ?? this.config.maxOutputTokens ?? 128_000,
+			temperature: reasoningOn ? undefined : 0,
+			system: [
+				supportsPromptCache
+					? {
+							text: systemPrompt,
+							type: "text",
+							cache_control: { type: "ephemeral" },
+						}
+					: { text: systemPrompt, type: "text" },
+			],
+			messages: anthropicMessages as Anthropic.MessageParam[],
+			stream: true,
+			tools: anthropicTools,
+			tool_choice: nativeToolsOn && !reasoningOn ? { type: "auto" } : undefined,
+		};
+
 		const stream = await client.messages.create(
-			{
-				model: model.id,
-				thinking: reasoningOn
-					? { type: "enabled", budget_tokens: budgetTokens }
-					: undefined,
-				max_tokens:
-					model.info.maxTokens ?? this.config.maxOutputTokens ?? 128_000,
-				temperature: reasoningOn ? undefined : 0,
-				system: supportsPromptCache
-					? [
-							{
-								text: systemPrompt,
-								type: "text",
-								cache_control: { type: "ephemeral" },
-							},
-						]
-					: [{ text: systemPrompt, type: "text" }],
-				messages: anthropicMessages as Anthropic.MessageParam[],
-				stream: true,
-				tools: anthropicTools,
-				tool_choice:
-					nativeToolsOn && !reasoningOn ? { type: "auto" } : undefined,
-			},
+			createParams as Anthropic.MessageCreateParamsStreaming,
 			requestOptions,
 		);
 
@@ -244,6 +244,7 @@ export class AnthropicHandler extends BaseHandler {
 						usageSnapshot.inputTokens,
 						usageSnapshot.outputTokens,
 						usageSnapshot.cacheReadTokens,
+						usageSnapshot.cacheWriteTokens,
 					),
 					id: responseId,
 				};
@@ -263,6 +264,7 @@ export class AnthropicHandler extends BaseHandler {
 						usageSnapshot.inputTokens,
 						usageSnapshot.outputTokens,
 						usageSnapshot.cacheReadTokens,
+						usageSnapshot.cacheWriteTokens,
 					),
 					id: responseId,
 				};

package/src/providers/handlers/base.test.ts

@@ -15,8 +15,14 @@ class TestHandler extends BaseHandler {
 		inputTokens: number,
 		outputTokens: number,
 		cacheReadTokens = 0,
+		cacheWriteTokens = 0,
 	): number | undefined {
-		return this.calculateCost(inputTokens, outputTokens, cacheReadTokens);
+		return this.calculateCost(
+			inputTokens,
+			outputTokens,
+			cacheReadTokens,
+			cacheWriteTokens,
+		);
 	}
 
 	public exposeAbortSignal(): AbortSignal {
@@ -45,7 +51,7 @@ describe("BaseHandler.calculateCost", () => {
 
 		const cost = handler.computeCost(1_000_000, 1_000_000, 100_000);
 
-		expect(cost).toBeCloseTo(17.73, 6);
+		expect(cost).toBeCloseTo(18.03, 6);
 	});
 });
 

package/src/providers/handlers/base.ts

@@ -10,6 +10,7 @@ import type {
 	ApiStream,
 	ApiStreamUsageChunk,
 	HandlerModelInfo,
+	ModelInfo,
 	ProviderConfig,
 } from "../types";
 import type { Message, ToolDefinition } from "../types/messages";
@@ -146,10 +147,26 @@ export abstract class BaseHandler implements ApiHandler {
 		});
 	}
 
+	protected supportsPromptCache(modelInfo?: ModelInfo): boolean {
+		const resolvedModelInfo =
+			modelInfo ??
+			this.config.modelInfo ??
+			this.config.knownModels?.[this.config.modelId];
+		const pricing = resolvedModelInfo?.pricing;
+
+		return (
+			resolvedModelInfo?.capabilities?.includes("prompt-cache") === true ||
+			this.config.capabilities?.includes("prompt-cache") === true ||
+			typeof pricing?.cacheRead === "number" ||
+			typeof pricing?.cacheWrite === "number"
+		);
+	}
+
 	protected calculateCost(
 		inputTokens: number,
 		outputTokens: number,
 		cacheReadTokens = 0,
+		cacheWriteTokens = 0,
 	): number | undefined {
 		const pricing = (
 			this.config.modelInfo ?? this.config.knownModels?.[this.config.modelId]
@@ -159,10 +176,14 @@ export abstract class BaseHandler implements ApiHandler {
 		}
 
 		return (
-			((inputTokens - cacheReadTokens) / 1_000_000) * pricing.input +
+			(inputTokens / 1_000_000) * pricing.input +
 			(outputTokens / 1_000_000) * pricing.output +
 			(cacheReadTokens > 0
 				? (cacheReadTokens / 1_000_000) * (pricing.cacheRead ?? 0)
+				: 0) +
+			(cacheWriteTokens > 0
+				? (cacheWriteTokens / 1_000_000) *
+					(pricing.cacheWrite ?? pricing.input * 1.25)
 				: 0)
 		);
 	}

package/src/providers/handlers/bedrock-base.ts

@@ -216,7 +216,7 @@ export class BedrockHandler extends BaseHandler {
 
 					yield {
 						type: "usage",
-						inputTokens: inputTokens - cacheReadTokens,
+						inputTokens,
 						outputTokens,
 						thoughtsTokenCount,
 						cacheReadTokens,
@@ -245,7 +245,7 @@ export class BedrockHandler extends BaseHandler {
 
 			yield {
 				type: "usage",
-				inputTokens: inputTokens - cacheReadTokens,
+				inputTokens,
 				outputTokens,
 				thoughtsTokenCount,
 				cacheReadTokens,

package/src/providers/handlers/community-sdk.test.ts

@@ -115,6 +115,39 @@ describe("Community SDK handlers", () => {
 			expect(usageChunk?.outputTokens).toBe(3);
 		});
 
+		it("keeps cached input tokens separate from total input tokens", async () => {
+			streamTextSpy.mockReturnValue({
+				fullStream: makeStreamParts([
+					{
+						type: "finish",
+						usage: { inputTokens: 10, outputTokens: 3, cachedInputTokens: 4 },
+					},
+				]),
+			});
+
+			const handler = new ClaudeCodeHandler({
+				providerId: "claude-code",
+				modelId: "sonnet",
+			});
+
+			const chunks: ApiStreamChunk[] = [];
+			for await (const chunk of handler.createMessage("System", [
+				{ role: "user", content: "Hi" },
+			])) {
+				chunks.push(chunk);
+			}
+
+			const usageChunk = chunks.find(
+				(chunk): chunk is Extract<ApiStreamChunk, { type: "usage" }> =>
+					chunk.type === "usage",
+			);
+			expect(usageChunk).toMatchObject({
+				inputTokens: 10,
+				outputTokens: 3,
+				cacheReadTokens: 4,
+			});
+		});
+
 		it("uses a fallback model id when model is missing", () => {
 			const handler = new ClaudeCodeHandler({
 				providerId: "claude-code",

package/src/providers/handlers/gemini-base.ts

@@ -18,7 +18,6 @@ import {
 import {
 	type ApiStream,
 	type HandlerModelInfo,
-	type ModelInfo,
 	type ProviderConfig,
 	supportsModelThinking,
 } from "../types";
@@ -258,7 +257,6 @@ export class GeminiHandler extends BaseHandler {
 
 			// Yield final usage
 			const totalCost = this.calculateGeminiCost(
-				info,
 				promptTokens,
 				outputTokens,
 				thoughtsTokenCount,
@@ -267,7 +265,7 @@ export class GeminiHandler extends BaseHandler {
 
 			yield {
 				type: "usage",
-				inputTokens: promptTokens - cacheReadTokens,
+				inputTokens: promptTokens,
 				outputTokens,
 				thoughtsTokenCount,
 				cacheReadTokens,
@@ -288,27 +286,16 @@ export class GeminiHandler extends BaseHandler {
 	}
 
 	private calculateGeminiCost(
-		info: ModelInfo,
 		inputTokens: number,
 		outputTokens: number,
 		thoughtsTokenCount: number,
 		cacheReadTokens: number,
 	): number | undefined {
-		const pricing = info.pricing;
-		if (!pricing?.input || !pricing?.output) {
-			return undefined;
-		}
-
-		const uncachedInputTokens = inputTokens - cacheReadTokens;
-		const inputCost = pricing.input * (uncachedInputTokens / 1_000_000);
-		const outputCost =
-			pricing.output * ((outputTokens + thoughtsTokenCount) / 1_000_000);
-		const cacheReadCost =
-			cacheReadTokens > 0
-				? (pricing.cacheRead ?? 0) * (cacheReadTokens / 1_000_000)
-				: 0;
-
-		return inputCost + outputCost + cacheReadCost;
+		return this.calculateCost(
+			inputTokens,
+			outputTokens + thoughtsTokenCount,
+			cacheReadTokens,
+		);
 	}
 }
 

package/src/providers/handlers/openai-base.ts

@@ -22,7 +22,6 @@ import type {
 	ModelInfo,
 	ProviderConfig,
 } from "../types";
-import { hasModelCapability } from "../types";
 import type { Message, ToolDefinition } from "../types/messages";
 import { retryStream } from "../utils/retry";
 import { ToolCallProcessor } from "../utils/tool-processor";
@@ -108,9 +107,7 @@ export class OpenAIBaseHandler extends BaseHandler {
 		messages: Message[],
 	): OpenAI.Chat.ChatCompletionMessageParam[] {
 		const model = this.getModel();
-		const supportsPromptCache =
-			hasModelCapability(model.info, "prompt-cache") ||
-			this.config.capabilities?.includes("prompt-cache") === true;
+		const supportsPromptCache = this.supportsPromptCache(model.info);
 		const systemMessage = supportsPromptCache
 			? ({
 					role: "system",
@@ -156,7 +153,8 @@ export class OpenAIBaseHandler extends BaseHandler {
 		const openAiMessages = this.getMessages(systemPrompt, messages);
 
 		// Build request options
-		const requestOptions: OpenAI.ChatCompletionCreateParamsStreaming = {
+		const requestOptions: Record<string, unknown> &
+			OpenAI.ChatCompletionCreateParamsStreaming = {
 			model: modelId,
 			messages: openAiMessages,
 			stream: true,
@@ -167,6 +165,17 @@ export class OpenAIBaseHandler extends BaseHandler {
 			}),
 		};
 
+		// Add top-level cache_control for OpenRouter with Anthropic models.
+		// This enables automatic caching where the cache breakpoint advances
+		// as the conversation grows, rather than relying on explicit per-block
+		// breakpoints which are limited to 4.
+		if (
+			this.config.providerId === "openrouter" &&
+			modelId.startsWith("anthropic/")
+		) {
+			requestOptions.cache_control = { type: "ephemeral" };
+		}
+
 		// Add max tokens if configured
 		const maxTokens = modelInfo.maxTokens ?? this.config.maxOutputTokens;
 		if (maxTokens) {
@@ -286,15 +295,16 @@ export class OpenAIBaseHandler extends BaseHandler {
 					cached_tokens?: number;
 					cache_write_tokens?: number;
 				};
-				prompt_cache_miss_tokens?: number;
 				cache_creation_input_tokens?: number;
 				cache_read_input_tokens?: number;
 			};
 			const cacheReadTokens =
-				usageWithCache.prompt_tokens_details?.cached_tokens ?? 0;
+				usageWithCache.prompt_tokens_details?.cached_tokens ??
+				usageWithCache.cache_read_input_tokens ??
+				0;
 			const cacheWriteTokens =
 				usageWithCache.prompt_tokens_details?.cache_write_tokens ??
-				usageWithCache.prompt_cache_miss_tokens ??
+				usageWithCache.cache_creation_input_tokens ??
 				0;
 
 			yield {
@@ -307,6 +317,7 @@ export class OpenAIBaseHandler extends BaseHandler {
 					inputTokens,
 					outputTokens,
 					cacheReadTokens,
+					cacheWriteTokens,
 				),
 				id: responseId,
 			};

package/src/providers/handlers/openai-responses.test.ts

@@ -210,4 +210,50 @@ describe("OpenAIResponsesHandler", () => {
 			},
 		});
 	});
+
+	it("keeps cached input tokens separate in usage chunks", () => {
+		const handler = new TestOpenAIResponsesHandler({
+			providerId: "openai-native",
+			modelId: "gpt-5.4",
+			apiKey: "test-key",
+			baseUrl: "https://example.com",
+			modelInfo: {
+				id: "gpt-5.4",
+				pricing: {
+					input: 1,
+					output: 2,
+					cacheRead: 0.5,
+				},
+			},
+		});
+
+		const chunks = handler.processChunkForTest({
+			type: "response.completed",
+			response: {
+				id: "resp_usage",
+				usage: {
+					input_tokens: 100,
+					output_tokens: 40,
+					input_tokens_details: {
+						cached_tokens: 25,
+					},
+					output_tokens_details: {
+						reasoning_tokens: 10,
+					},
+				},
+			},
+		});
+
+		expect(chunks[0]).toMatchObject({
+			type: "usage",
+			inputTokens: 100,
+			outputTokens: 40,
+			cacheReadTokens: 25,
+			cacheWriteTokens: 0,
+		});
+		expect(chunks[0]?.type).toBe("usage");
+		if (chunks[0]?.type === "usage") {
+			expect(chunks[0].totalCost).toBeCloseTo(0.0001925, 10);
+		}
+	});
 });

package/src/providers/handlers/openai-responses.ts

@@ -565,23 +565,19 @@ export class OpenAIResponsesHandler extends BaseHandler {
 					const inputTokens = usage.input_tokens || 0;
 					const outputTokens = usage.output_tokens || 0;
 					const cacheReadTokens =
-						usage.output_tokens_details?.reasoning_tokens || 0;
-					const cacheWriteTokens =
 						usage.input_tokens_details?.cached_tokens || 0;
+					const cacheWriteTokens = 0;
 
 					const totalCost = this.calculateCost(
 						inputTokens,
 						outputTokens,
 						cacheReadTokens,
-					);
-					const nonCachedInputTokens = Math.max(
-						0,
-						inputTokens - cacheReadTokens - cacheWriteTokens,
+						cacheWriteTokens,
 					);
 
 					yield {
 						type: "usage",
-						inputTokens: nonCachedInputTokens,
+						inputTokens,
 						outputTokens,
 						cacheWriteTokens,
 						cacheReadTokens,

package/src/providers/handlers/r1-base.ts

@@ -255,19 +255,18 @@ export class R1BaseHandler extends BaseHandler {
 		const cacheReadTokens = r1Usage.prompt_cache_hit_tokens ?? 0;
 		const cacheWriteTokens = r1Usage.prompt_cache_miss_tokens ?? 0;
 
-		// Calculate non-cached input tokens (will always be 0 for DeepSeek since input = read + write)
-		const nonCachedInputTokens = Math.max(
-			0,
-			inputTokens - cacheReadTokens - cacheWriteTokens,
-		);
-
 		yield {
 			type: "usage",
-			inputTokens: nonCachedInputTokens,
+			inputTokens,
 			outputTokens,
 			cacheReadTokens,
 			cacheWriteTokens,
-			totalCost: this.calculateCost(inputTokens, outputTokens, cacheReadTokens),
+			totalCost: this.calculateCost(
+				inputTokens,
+				outputTokens,
+				cacheReadTokens,
+				cacheWriteTokens,
+			),
 			id: responseId,
 		};
 	}

package/src/providers/handlers/vertex.ts

@@ -189,7 +189,7 @@ export class VertexHandler extends BaseHandler {
 		if (!isClaudeModel(model.id)) {
 			return this.ensureGeminiHandler().getMessages(systemPrompt, messages);
 		}
-		const supportsPromptCache = hasModelCapability(model.info, "prompt-cache");
+		const supportsPromptCache = this.supportsPromptCache(model.info);
 		return convertToAnthropicMessages(messages, supportsPromptCache);
 	}
 
@@ -226,7 +226,7 @@ export class VertexHandler extends BaseHandler {
 		const budgetTokens = this.config.thinkingBudgetTokens ?? 0;
 		const reasoningOn =
 			hasModelCapability(model.info, "reasoning") && budgetTokens > 0;
-		const promptCacheOn = hasModelCapability(model.info, "prompt-cache");
+		const promptCacheOn = this.supportsPromptCache(model.info);
 
 		const providerOptions: Record<string, unknown> = {};
 		if (reasoningOn) {
@@ -251,8 +251,18 @@ export class VertexHandler extends BaseHandler {
 		yield* emitAiSdkStream(stream, {
 			responseId,
 			errorMessage: "Vertex Anthropic stream failed",
-			calculateCost: (inputTokens, outputTokens, cacheReadTokens) =>
-				this.calculateCost(inputTokens, outputTokens, cacheReadTokens),
+			calculateCost: (
+				inputTokens,
+				outputTokens,
+				cacheReadTokens,
+				cacheWriteTokens,
+			) =>
+				this.calculateCost(
+					inputTokens,
+					outputTokens,
+					cacheReadTokens,
+					cacheWriteTokens,
+				),
 			reasoningTypes: ["reasoning-delta"],
 			enableToolCalls: true,
 			toolCallArgsOrder: ["input", "args"],