deepinfra/claude-3.7-sonnet-latest (#5062)

* integrate deepinfra/claude-3.7-sonnet;

* update snapshot

Author: juliettech13

packages/__tests__/cost/__snapshots__/registrySnapshots.test.ts.snap

@@ -478,6 +478,26 @@ exports[`Registry Snapshots endpoint configurations snapshot 1`] = `
         "us-east-1",
       ],
     },
+    "claude-3.7-sonnet:deepinfra": {
+      "context": 200000,
+      "crossRegion": false,
+      "maxTokens": 16384,
+      "modelId": "anthropic/claude-3-7-sonnet-latest",
+      "parameters": [
+        "max_tokens",
+        "stop",
+        "temperature",
+        "tool_choice",
+        "tools",
+        "top_k",
+        "top_p",
+      ],
+      "provider": "deepinfra",
+      "ptbEnabled": true,
+      "regions": [
+        "*",
+      ],
+    },
     "claude-3.7-sonnet:openrouter": {
       "context": 200000,
       "crossRegion": false,
@@ -3645,6 +3665,7 @@ exports[`Registry Snapshots model coverage snapshot 1`] = `
   "anthropic/claude-3.7-sonnet": [
     "anthropic",
     "bedrock",
+    "deepinfra",
     "openrouter",
     "vertex",
   ],
@@ -4004,6 +4025,16 @@ exports[`Registry Snapshots pricing snapshot 1`] = `
         "web_search": 0.00001,
       },
     ],
+    "deepinfra": [
+      {
+        "cacheMultipliers": {
+          "cachedInput": 0.1,
+        },
+        "input": 0.0000033,
+        "output": 0.0000165,
+        "threshold": 0,
+      },
+    ],
     "openrouter": [
       {
         "input": 0.000003165,
@@ -4889,6 +4920,7 @@ exports[`Registry Snapshots verify registry state 1`] = `
       "providers": [
         "anthropic",
         "bedrock",
+        "deepinfra",
         "openrouter",
         "vertex",
       ],

packages/cost/models/authors/anthropic/claude-3.7-sonnet/endpoints.ts

@@ -21,8 +21,8 @@ export const endpoints = {
         },
       },
     ],
-    contextLength: 200000,
-    maxCompletionTokens: 64000,
+    contextLength: 200_000,
+    maxCompletionTokens: 64_000,
     supportedParameters: [
       "tools",
       "tool_choice",
@@ -58,8 +58,8 @@ export const endpoints = {
         },
       },
     ],
-    contextLength: 200000,
-    maxCompletionTokens: 64000,
+    contextLength: 200_000,
+    maxCompletionTokens: 64_000,
     supportedParameters: [
       "tools",
       "tool_choice",
@@ -93,8 +93,8 @@ export const endpoints = {
         },
       },
     ],
-    contextLength: 200000,
-    maxCompletionTokens: 64000,
+    contextLength: 200_000,
+    maxCompletionTokens: 64_000,
     supportedParameters: [
       "tools",
       "tool_choice",
@@ -121,8 +121,8 @@ export const endpoints = {
         output: 0.00001583, // $15.83/1M - worst-case: $15.00/1M (Google) * 1.055,
       },
     ],
-    contextLength: 200000,
-    maxCompletionTokens: 64000,
+    contextLength: 200_000,
+    maxCompletionTokens: 64_000,
     supportedParameters: [
       "max_tokens",
       "temperature",
@@ -138,6 +138,36 @@ export const endpoints = {
       "*": {},
     },
   },
+  "claude-3.7-sonnet:deepinfra": {
+    provider: "deepinfra",
+    author: "anthropic",
+    providerModelId: "anthropic/claude-3-7-sonnet-latest",
+    pricing: [
+      {
+        threshold: 0,
+        input: 0.0000033, // $3.30/1M
+        output: 0.0000165, // $16.50/1M
+        cacheMultipliers: {
+          cachedInput: 0.1, // $0.33/1M
+        },
+      },
+    ],
+    contextLength: 200_000,
+    maxCompletionTokens: 16_384,
+    supportedParameters: [
+      "max_tokens",
+      "temperature",
+      "top_p",
+      "top_k",
+      "stop",
+      "tools",
+      "tool_choice",
+    ],
+    ptbEnabled: true,
+    endpointConfigs: {
+      "*": {},
+    },
+  },
 } satisfies Partial<
   Record<`${Claude37SonnetModelName}:${ModelProviderName}`, ModelProviderConfig>
 >;

worker/test/ai-gateway/registry-anthropic.spec.ts

@@ -26,6 +26,12 @@ const bedrockAuthExpectations = {
   },
 };
 
+const deepinfraAuthExpectations = {
+  headers: {
+    Authorization: /^Bearer /,
+  },
+};
+
 describe("Anthropic Registry Tests", () => {
   beforeEach(() => {
     // Clear all mocks between tests
@@ -278,6 +284,23 @@ describe("Anthropic Registry Tests", () => {
           },
         }));
 
+      it("should handle deepinfra provider", () =>
+        runGatewayTest({
+          model: "claude-3.7-sonnet/deepinfra",
+          expected: {
+            providers: [
+              {
+                url: "https://api.deepinfra.com/v1/openai/chat/completions",
+                response: "success",
+                model: "anthropic/claude-3-7-sonnet-latest",
+                data: createAnthropicMockResponse("claude-3.7-sonnet"),
+                expects: deepinfraAuthExpectations,
+              },
+            ],
+            finalStatus: 200,
+          },
+        }));
+
       it("should auto-select provider when none specified", () =>
         runGatewayTest({
           model: "claude-3.7-sonnet",
@@ -664,6 +687,54 @@ describe("Anthropic Registry Tests", () => {
             finalStatus: 401,
           },
         }));
+
+      it("should handle DeepInfra provider failure for claude-3.7-sonnet", () =>
+        runGatewayTest({
+          model: "claude-3.7-sonnet/deepinfra",
+          expected: {
+            providers: [
+              {
+                url: "https://api.deepinfra.com/v1/openai/chat/completions",
+                response: "failure",
+                statusCode: 500,
+                errorMessage: "DeepInfra service unavailable",
+              },
+            ],
+            finalStatus: 500,
+          },
+        }));
+
+      it("should handle rate limiting from DeepInfra for claude-3.7-sonnet", () =>
+        runGatewayTest({
+          model: "claude-3.7-sonnet/deepinfra",
+          expected: {
+            providers: [
+              {
+                url: "https://api.deepinfra.com/v1/openai/chat/completions",
+                response: "failure",
+                statusCode: 429,
+                errorMessage: "Rate limit exceeded",
+              },
+            ],
+            finalStatus: 429,
+          },
+        }));
+
+      it("should handle authentication failure from DeepInfra for claude-3.7-sonnet", () =>
+        runGatewayTest({
+          model: "claude-3.7-sonnet/deepinfra",
+          expected: {
+            providers: [
+              {
+                url: "https://api.deepinfra.com/v1/openai/chat/completions",
+                response: "failure",
+                statusCode: 401,
+                errorMessage: "Invalid API key",
+              },
+            ],
+            finalStatus: 401,
+          },
+        }));
     });
   });
 });