Add getOpenRouterModelEndpoints

Author: cte

src/api/providers/fetchers/__tests__/fixtures/openrouter-model-endpoints.json

@@ -0,0 +1,25 @@
+[
+	{
+		"scope": "https://openrouter.ai:443",
+		"method": "GET",
+		"path": "/api/v1/models/google/gemini-2.5-pro-preview/endpoints",
+		"body": "",
+		"status": 200,
+		"response": [
+			"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"
+		],
+		"rawHeaders": {
+			"access-control-allow-origin": "*",
+			"cache-control": "s-maxage=300, stale-while-revalidate=600",
+			"cf-ray": "93ed496b8e0a0fb1-LAX",
+			"connection": "close",
+			"content-encoding": "br",
+			"content-type": "application/json",
+			"date": "Mon, 12 May 2025 22:17:32 GMT",
+			"server": "cloudflare",
+			"transfer-encoding": "chunked",
+			"vary": "Accept-Encoding"
+		},
+		"responseIsBinary": false
+	}
+]

src/api/providers/fetchers/__tests__/fixtures/openrouter-models.json

@@ -6,15 +6,15 @@ import { back as nockBack } from "nock"
 
 import { PROMPT_CACHING_MODELS } from "../../../../shared/api"
 
-import { getOpenRouterModels } from "../openrouter"
+import { getOpenRouterModelEndpoints, getOpenRouterModels } from "../openrouter"
 
 nockBack.fixtures = path.join(__dirname, "fixtures")
 nockBack.setMode("lockdown")
 
-describe("OpenRouter API", () => {
+describe.skip("OpenRouter API", () => {
 	describe("getOpenRouterModels", () => {
 		// This flakes in CI (probably related to Nock). Need to figure out why.
-		it.skip("fetches models and validates schema", async () => {
+		it("fetches models and validates schema", async () => {
 			const { nockDone } = await nockBack("openrouter-models.json")
 
 			const models = await getOpenRouterModels()
@@ -95,4 +95,40 @@ describe("OpenRouter API", () => {
 			nockDone()
 		})
 	})
+
+	describe("getOpenRouterModelEndpoints", () => {
+		it("fetches model endpoints and validates schema", async () => {
+			const { nockDone } = await nockBack("openrouter-model-endpoints.json")
+			const endpoints = await getOpenRouterModelEndpoints("google/gemini-2.5-pro-preview")
+
+			expect(endpoints).toEqual({
+				Google: {
+					maxTokens: 0,
+					contextWindow: 1048576,
+					supportsImages: true,
+					supportsPromptCache: true,
+					inputPrice: 1.25,
+					outputPrice: 10,
+					cacheWritesPrice: 1.625,
+					cacheReadsPrice: 0.31,
+					description: undefined,
+					thinking: false,
+				},
+				"Google AI Studio": {
+					maxTokens: 0,
+					contextWindow: 1048576,
+					supportsImages: true,
+					supportsPromptCache: true,
+					inputPrice: 1.25,
+					outputPrice: 10,
+					cacheWritesPrice: 1.625,
+					cacheReadsPrice: 0.31,
+					description: undefined,
+					thinking: false,
+				},
+			})
+
+			nockDone()
+		})
+	})
 })

src/api/providers/fetchers/__tests__/openrouter.test.ts

@@ -4,97 +4,106 @@ import { z } from "zod"
 import { ApiHandlerOptions, ModelInfo, anthropicModels, COMPUTER_USE_MODELS } from "../../../shared/api"
 import { parseApiPrice } from "../../../utils/cost"
 
-// https://openrouter.ai/api/v1/models
-export const openRouterModelSchema = z.object({
-	id: z.string(),
+/**
+ * OpenRouterBaseModel
+ */
+
+const openRouterArchitectureSchema = z.object({
+	modality: z.string().nullish(),
+	tokenizer: z.string().nullish(),
+})
+
+const openRouterPricingSchema = z.object({
+	prompt: z.string().nullish(),
+	completion: z.string().nullish(),
+	input_cache_write: z.string().nullish(),
+	input_cache_read: z.string().nullish(),
+})
+
+const modelRouterBaseModelSchema = z.object({
 	name: z.string(),
 	description: z.string().optional(),
 	context_length: z.number(),
 	max_completion_tokens: z.number().nullish(),
-	architecture: z
-		.object({
-			modality: z.string().nullish(),
-			tokenizer: z.string().nullish(),
-		})
-		.optional(),
-	pricing: z
-		.object({
-			prompt: z.string().nullish(),
-			completion: z.string().nullish(),
-			input_cache_write: z.string().nullish(),
-			input_cache_read: z.string().nullish(),
-		})
-		.optional(),
-	top_provider: z
-		.object({
-			max_completion_tokens: z.number().nullish(),
-		})
-		.optional(),
+	pricing: openRouterPricingSchema.optional(),
+})
+
+export type OpenRouterBaseModel = z.infer<typeof modelRouterBaseModelSchema>
+
+/**
+ * OpenRouterModel
+ */
+
+export const openRouterModelSchema = modelRouterBaseModelSchema.extend({
+	id: z.string(),
+	architecture: openRouterArchitectureSchema.optional(),
+	top_provider: z.object({ max_completion_tokens: z.number().nullish() }).optional(),
 })
 
 export type OpenRouterModel = z.infer<typeof openRouterModelSchema>
 
+/**
+ * OpenRouterModelEndpoint
+ */
+
+export const openRouterModelEndpointSchema = modelRouterBaseModelSchema.extend({
+	provider_name: z.string(),
+})
+
+export type OpenRouterModelEndpoint = z.infer<typeof openRouterModelEndpointSchema>
+
+/**
+ * OpenRouterModelsResponse
+ */
+
 const openRouterModelsResponseSchema = z.object({
 	data: z.array(openRouterModelSchema),
 })
 
 type OpenRouterModelsResponse = z.infer<typeof openRouterModelsResponseSchema>
 
+/**
+ * OpenRouterModelEndpointsResponse
+ */
+
+const openRouterModelEndpointsResponseSchema = z.object({
+	data: z.object({
+		id: z.string(),
+		name: z.string(),
+		description: z.string().optional(),
+		architecture: openRouterArchitectureSchema.optional(),
+		endpoints: z.array(openRouterModelEndpointSchema),
+	}),
+})
+
+type OpenRouterModelEndpointsResponse = z.infer<typeof openRouterModelEndpointsResponseSchema>
+
+/**
+ * getOpenRouterModels
+ */
+
 export async function getOpenRouterModels(options?: ApiHandlerOptions): Promise<Record<string, ModelInfo>> {
 	const models: Record<string, ModelInfo> = {}
 	const baseURL = options?.openRouterBaseUrl || "https://openrouter.ai/api/v1"
 
 	try {
 		const response = await axios.get<OpenRouterModelsResponse>(`${baseURL}/models`)
 		const result = openRouterModelsResponseSchema.safeParse(response.data)
-		const rawModels = result.success ? result.data.data : response.data.data
+		const data = result.success ? result.data.data : response.data.data
 
 		if (!result.success) {
 			console.error("OpenRouter models response is invalid", result.error.format())
 		}
 
-		for (const rawModel of rawModels) {
-			const cacheWritesPrice = rawModel.pricing?.input_cache_write
-				? parseApiPrice(rawModel.pricing?.input_cache_write)
-				: undefined
-
-			const cacheReadsPrice = rawModel.pricing?.input_cache_read
-				? parseApiPrice(rawModel.pricing?.input_cache_read)
-				: undefined
-
-			const supportsPromptCache =
-				typeof cacheWritesPrice !== "undefined" && typeof cacheReadsPrice !== "undefined"
-
-			const modelInfo: ModelInfo = {
-				maxTokens: rawModel.id.startsWith("anthropic/") ? rawModel.top_provider?.max_completion_tokens : 0,
-				contextWindow: rawModel.context_length,
-				supportsImages: rawModel.architecture?.modality?.includes("image"),
-				supportsPromptCache,
-				inputPrice: parseApiPrice(rawModel.pricing?.prompt),
-				outputPrice: parseApiPrice(rawModel.pricing?.completion),
-				cacheWritesPrice,
-				cacheReadsPrice,
-				description: rawModel.description,
-				thinking: rawModel.id === "anthropic/claude-3.7-sonnet:thinking",
-			}
-
-			// The OpenRouter model definition doesn't give us any hints about
-			// computer use, so we need to set that manually.
-			if (COMPUTER_USE_MODELS.has(rawModel.id)) {
-				modelInfo.supportsComputerUse = true
-			}
-
-			// Claude 3.7 Sonnet is a "hybrid" thinking model, and the `maxTokens`
-			// values can be configured. For the non-thinking variant we want to
-			// use 8k. The `thinking` variant can be run in 64k and 128k modes,
-			// and we want to use 128k.
-			if (rawModel.id.startsWith("anthropic/claude-3.7-sonnet")) {
-				modelInfo.maxTokens = rawModel.id.includes("thinking")
-					? anthropicModels["claude-3-7-sonnet-20250219:thinking"].maxTokens
-					: anthropicModels["claude-3-7-sonnet-20250219"].maxTokens
-			}
-
-			models[rawModel.id] = modelInfo
+		for (const model of data) {
+			const { id, architecture, top_provider } = model
+
+			models[id] = parseOpenRouterModel({
+				id,
+				model,
+				modality: architecture?.modality,
+				maxTokens: id.startsWith("anthropic/") ? top_provider?.max_completion_tokens : 0,
+			})
 		}
 	} catch (error) {
 		console.error(
@@ -104,3 +113,97 @@ export async function getOpenRouterModels(options?: ApiHandlerOptions): Promise<
 
 	return models
 }
+
+/**
+ * getOpenRouterModelEndpoints
+ */
+
+export async function getOpenRouterModelEndpoints(
+	modelId: string,
+	options?: ApiHandlerOptions,
+): Promise<Record<string, ModelInfo>> {
+	const models: Record<string, ModelInfo> = {}
+	const baseURL = options?.openRouterBaseUrl || "https://openrouter.ai/api/v1"
+
+	try {
+		const response = await axios.get<OpenRouterModelEndpointsResponse>(`${baseURL}/models/${modelId}/endpoints`)
+		const result = openRouterModelEndpointsResponseSchema.safeParse(response.data)
+		const data = result.success ? result.data.data : response.data.data
+
+		if (!result.success) {
+			console.error("OpenRouter model endpoints response is invalid", result.error.format())
+		}
+
+		const { id, architecture, endpoints } = data
+
+		for (const endpoint of endpoints) {
+			models[endpoint.provider_name] = parseOpenRouterModel({
+				id,
+				model: endpoint,
+				modality: architecture?.modality,
+				maxTokens: id.startsWith("anthropic/") ? endpoint.max_completion_tokens : 0,
+			})
+		}
+	} catch (error) {
+		console.error(
+			`Error fetching OpenRouter model endpoints: ${JSON.stringify(error, Object.getOwnPropertyNames(error), 2)}`,
+		)
+	}
+
+	return models
+}
+
+/**
+ * parseOpenRouterModel
+ */
+
+export const parseOpenRouterModel = ({
+	id,
+	model,
+	modality,
+	maxTokens,
+}: {
+	id: string
+	model: OpenRouterBaseModel
+	modality: string | null | undefined
+	maxTokens: number | null | undefined
+}): ModelInfo => {
+	const cacheWritesPrice = model.pricing?.input_cache_write
+		? parseApiPrice(model.pricing?.input_cache_write)
+		: undefined
+
+	const cacheReadsPrice = model.pricing?.input_cache_read ? parseApiPrice(model.pricing?.input_cache_read) : undefined
+
+	const supportsPromptCache = typeof cacheWritesPrice !== "undefined" && typeof cacheReadsPrice !== "undefined"
+
+	const modelInfo: ModelInfo = {
+		maxTokens: maxTokens || 0,
+		contextWindow: model.context_length,
+		supportsImages: modality?.includes("image") ?? false,
+		supportsPromptCache,
+		inputPrice: parseApiPrice(model.pricing?.prompt),
+		outputPrice: parseApiPrice(model.pricing?.completion),
+		cacheWritesPrice,
+		cacheReadsPrice,
+		description: model.description,
+		thinking: id === "anthropic/claude-3.7-sonnet:thinking",
+	}
+
+	// The OpenRouter model definition doesn't give us any hints about
+	// computer use, so we need to set that manually.
+	if (COMPUTER_USE_MODELS.has(id)) {
+		modelInfo.supportsComputerUse = true
+	}
+
+	// Claude 3.7 Sonnet is a "hybrid" thinking model, and the `maxTokens`
+	// values can be configured. For the non-thinking variant we want to
+	// use 8k. The `thinking` variant can be run in 64k and 128k modes,
+	// and we want to use 128k.
+	if (id.startsWith("anthropic/claude-3.7-sonnet")) {
+		modelInfo.maxTokens = id.includes("thinking")
+			? anthropicModels["claude-3-7-sonnet-20250219:thinking"].maxTokens
+			: anthropicModels["claude-3-7-sonnet-20250219"].maxTokens
+	}
+
+	return modelInfo
+}