diff --git a/docs/field-conversions.md b/docs/field-conversions.md
index c952521..09fc7c0 100644
--- a/docs/field-conversions.md
+++ b/docs/field-conversions.md
@@ -188,9 +188,10 @@ Complete mapping of all fields across the 12 pairwise converters between OpenAI
 | `choices[0].message.annotations` | `content[]` web_search_tool_result | |
 | `choices[0].message.tool_calls` | `content[]` tool_use | |
 | `finish_reason` | `stop_reason` | stop→end_turn, tool_calls→tool_use, length→max_tokens, content_filter→refusal |
-| `usage.prompt_tokens` | `usage.input_tokens` | |
+| `usage.prompt_tokens` - (`cached_tokens` + `cache_write_tokens`) | `usage.input_tokens` | |
 | `usage.completion_tokens` | `usage.output_tokens` | |
 | `prompt_tokens_details.cached_tokens` | `cache_read_input_tokens` | |
+| `prompt_tokens_details.cache_write_tokens` | `cache_creation_input_tokens` | |
 | `prompt_tokens_details.web_search` | `server_tool_use.web_search_requests` | |
 
 Usage also re-emits enriched OpenAI-style fields (`prompt_tokens`, `completion_tokens`, `total_tokens`, `prompt_tokens_details`, `completion_tokens_details`, audio token fields); origin `*_details` objects are merged in to preserve extra fields. Streaming reuses the same `buildUsage`, so `message_delta` carries real `output_tokens` from the trailing usage chunk.
@@ -233,7 +234,10 @@ Usage also re-emits enriched OpenAI-style fields (`prompt_tokens`, `completion_t
 | `output[]` function_call | `content[]` tool_use | |
 | `output[]` message | `content[]` text | |
 | `status` | `stop_reason` | completed→end_turn/tool_use, incomplete→max_tokens, failed→end_turn |
-| `usage.*` | `usage.*` | cached_tokens→cache_read_input_tokens |
+| `usage.input_tokens` - (`cached_tokens` + `cache_write_tokens`) | `usage.input_tokens` | |
+| `usage.input_tokens_details.cached_tokens` | `usage.cache_read_input_tokens` | |
+| `usage.input_tokens_details.cache_write_tokens` | `usage.cache_creation_input_tokens` | |
+| Other `usage.*` | `usage.*` | Reasoning and enriched OpenAI-style fields preserved |
 
 ### Stream: text ✅ tool_calls ✅ thinking ✅ usage ✅
 
@@ -273,7 +277,7 @@ Usage also re-emits enriched OpenAI-style fields (`prompt_tokens`, `completion_t
 | `candidates[0].content.parts[]` text | `content[]` text | |
 | `candidates[0].content.parts[]` functionCall | `content[]` tool_use | |
 | `candidates[0].finishReason` | `stop_reason` | STOP→end_turn, MAX_TOKENS→max_tokens, SAFETY→refusal |
-| `usageMetadata.promptTokenCount` + `toolUsePromptTokenCount` | `usage.input_tokens` | Combined |
+| `usageMetadata.promptTokenCount` + `toolUsePromptTokenCount` - `cachedContentTokenCount` | `usage.input_tokens` | Combined |
 | `usageMetadata.candidatesTokenCount` + `thoughtsTokenCount` | `usage.output_tokens` | Combined |
 | `usageMetadata.cachedContentTokenCount` | `cache_read_input_tokens` | |
 
diff --git a/package.json b/package.json
index a397e73..117ac7c 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@zenmux/rosetta-ai",
-  "version": "1.0.16",
+  "version": "1.0.17",
   "description": "Universal translator between AI provider protocols",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -60,4 +60,4 @@
     "typescript": "^6.0.2",
     "typescript-eslint": "^8.58.2"
   }
-}
\ No newline at end of file
+}
diff --git a/src/messages/__tests__/chat-completions.test.ts b/src/messages/__tests__/chat-completions.test.ts
index 9d0bb5b..0083be5 100644
--- a/src/messages/__tests__/chat-completions.test.ts
+++ b/src/messages/__tests__/chat-completions.test.ts
@@ -972,7 +972,7 @@ describe("MessagesToChatCompletionConverter", () => {
       expect(converter.convertResponse(input).stop_reason).toBe("tool_use");
     });
 
-    it("converts usage with cache details", () => {
+    it("separates cache read and write tokens from input_tokens", () => {
       const input: OpenAI.ChatCompletion = {
         id: "id",
         object: "chat.completion",
@@ -990,13 +990,16 @@ describe("MessagesToChatCompletionConverter", () => {
           prompt_tokens: 100,
           completion_tokens: 50,
           total_tokens: 150,
-          prompt_tokens_details: { cached_tokens: 30 },
-        },
+          prompt_tokens_details: { cached_tokens: 30, cache_write_tokens: 20 },
+        } as any,
       };
       const result = converter.convertResponse(input);
-      expect(result.usage.input_tokens).toBe(100);
+      expect(result.usage.input_tokens).toBe(50);
       expect(result.usage.output_tokens).toBe(50);
       expect(result.usage.cache_read_input_tokens).toBe(30);
+      expect(result.usage.cache_creation_input_tokens).toBe(20);
+      expect((result.usage as any).prompt_tokens).toBe(100);
+      expect((result.usage as any).total_tokens).toBe(150);
     });
 
     it("enriches usage with OpenAI-style fields", () => {
@@ -1030,6 +1033,7 @@ describe("MessagesToChatCompletionConverter", () => {
       expect(usage.prompt_tokens_details).toEqual({ cached_tokens: 30, audio_tokens: 4 });
       expect(usage.audio_input_tokens).toBe(4);
       expect(usage.service_tier).toBe("standard");
+      expect(usage.input_tokens).toBe(70);
       expect(usage.cache_creation_input_tokens).toBe(0);
     });
 
@@ -1491,7 +1495,12 @@ describe("MessagesToChatCompletionConverter", () => {
             created: 1700000000,
             model: "gpt-4o",
             choices: [],
-            usage: { prompt_tokens: 12, completion_tokens: 7, total_tokens: 19 },
+            usage: {
+              prompt_tokens: 12,
+              completion_tokens: 7,
+              total_tokens: 19,
+              prompt_tokens_details: { cached_tokens: 3, cache_write_tokens: 2 },
+            },
           } as unknown as OpenAI.ChatCompletionChunk,
         ];
 
@@ -1502,7 +1511,9 @@ describe("MessagesToChatCompletionConverter", () => {
 
         expect(msgDelta).toBeDefined();
         expect(msgDelta.usage.output_tokens).toBe(7);
-        expect(msgDelta.usage.input_tokens).toBe(12);
+        expect(msgDelta.usage.input_tokens).toBe(7);
+        expect(msgDelta.usage.cache_read_input_tokens).toBe(3);
+        expect(msgDelta.usage.cache_creation_input_tokens).toBe(2);
         expect((msgDelta.usage as any).completion_tokens).toBe(7);
         expect((msgDelta.usage as any).prompt_tokens).toBe(12);
         expect((msgDelta.usage as any).total_tokens).toBe(19);
diff --git a/src/messages/__tests__/gemini.test.ts b/src/messages/__tests__/gemini.test.ts
index 588183b..eae5d79 100644
--- a/src/messages/__tests__/gemini.test.ts
+++ b/src/messages/__tests__/gemini.test.ts
@@ -693,6 +693,9 @@ describe("MessagesToGeminiConverter", () => {
       expect(usage.prompt_tokens).toBe(110);
       expect(usage.completion_tokens).toBe(70);
       expect(usage.total_tokens).toBe(180);
+      expect(usage.input_tokens).toBe(80);
+      expect(usage.cache_read_input_tokens).toBe(30);
+      expect(usage.cache_creation_input_tokens).toBe(0);
       expect(usage.prompt_tokens_details.cached_tokens).toBe(30);
       expect(usage.completion_tokens_details.reasoning_tokens).toBe(20);
       expect(usage.tool_use).toBe(10);
@@ -945,6 +948,7 @@ describe("MessagesToGeminiConverter", () => {
             promptTokenCount: 10,
             candidatesTokenCount: 5,
             totalTokenCount: 15,
+            cachedContentTokenCount: 3,
           } as any,
         })
       );
@@ -952,6 +956,8 @@ describe("MessagesToGeminiConverter", () => {
       const msgDelta = events.find(e => e.type === "message_delta") as any;
       expect(msgDelta).toBeDefined();
       expect(msgDelta.delta.stop_reason).toBe("end_turn");
+      expect(msgDelta.usage.input_tokens).toBe(7);
+      expect(msgDelta.usage.cache_read_input_tokens).toBe(3);
 
       const msgStop = events.find(e => e.type === "message_stop");
       expect(msgStop).toBeDefined();
diff --git a/src/messages/__tests__/responses.test.ts b/src/messages/__tests__/responses.test.ts
index 8595ac9..8e96738 100644
--- a/src/messages/__tests__/responses.test.ts
+++ b/src/messages/__tests__/responses.test.ts
@@ -341,22 +341,25 @@ describe("MessagesToResponsesConverter", () => {
       expect(result.stop_reason).toBe("max_tokens");
     });
 
-    it("includes cache tokens in usage", () => {
+    it("separates cache read and write tokens from input_tokens", () => {
       const result = converter.convertResponse(
         makeResponse({
           usage: {
             input_tokens: 100,
             output_tokens: 50,
             total_tokens: 150,
-            input_tokens_details: { cached_tokens: 30 },
+            input_tokens_details: { cached_tokens: 30, cache_write_tokens: 20 },
             output_tokens_details: { reasoning_tokens: 0 },
-          },
+          } as any,
         })
       );
 
-      expect(result.usage.input_tokens).toBe(100);
+      expect(result.usage.input_tokens).toBe(50);
       expect(result.usage.output_tokens).toBe(50);
       expect(result.usage.cache_read_input_tokens).toBe(30);
+      expect(result.usage.cache_creation_input_tokens).toBe(20);
+      expect((result.usage as any).prompt_tokens).toBe(100);
+      expect((result.usage as any).total_tokens).toBe(150);
     });
 
     it("counts only completed web search calls", () => {
@@ -651,7 +654,7 @@ describe("MessagesToResponsesConverter", () => {
             input_tokens: 10,
             output_tokens: 5,
             total_tokens: 15,
-            input_tokens_details: { cached_tokens: 3 },
+            input_tokens_details: { cached_tokens: 3, cache_write_tokens: 2 },
             output_tokens_details: { reasoning_tokens: 0 },
           },
         } as any,
@@ -661,7 +664,10 @@ describe("MessagesToResponsesConverter", () => {
       const msgDelta = events.find(
         e => e.type === "message_delta"
       ) as Anthropic.RawMessageDeltaEvent;
+      expect(msgDelta.usage.input_tokens).toBe(5);
       expect(msgDelta.usage.output_tokens).toBe(5);
+      expect(msgDelta.usage.cache_read_input_tokens).toBe(3);
+      expect(msgDelta.usage.cache_creation_input_tokens).toBe(2);
     });
 
     it("emits tool_use stop_reason when function_call is in output", () => {
diff --git a/src/messages/chat-completions.ts b/src/messages/chat-completions.ts
index 85347a9..8187a65 100644
--- a/src/messages/chat-completions.ts
+++ b/src/messages/chat-completions.ts
@@ -183,6 +183,8 @@ export class MessagesToChatCompletionConverter {
     const outputTokens = u.completion_tokens ?? 0;
     const totalTokens = u.total_tokens ?? inputTokens + outputTokens;
     const cachedTokens = u.prompt_tokens_details?.cached_tokens ?? 0;
+    const cacheWriteTokens = (u.prompt_tokens_details as any)?.cache_write_tokens ?? 0;
+    const uncachedInputTokens = Math.max(0, inputTokens - cachedTokens - cacheWriteTokens);
     const reasoningTokens = u.completion_tokens_details?.reasoning_tokens ?? 0;
     const webSearch: number = (u.prompt_tokens_details as any)?.web_search ?? 0;
 
@@ -199,7 +201,7 @@ export class MessagesToChatCompletionConverter {
         cached_tokens: cachedTokens,
         ...(webSearch > 0 ? { web_search: webSearch } : {}),
       },
-      input_tokens: inputTokens,
+      input_tokens: uncachedInputTokens,
       output_tokens: outputTokens,
       cache_read_input_tokens: cachedTokens,
       server_tool_use:
@@ -209,7 +211,7 @@ export class MessagesToChatCompletionConverter {
               web_search_requests: webSearch,
             }
           : null,
-      cache_creation_input_tokens: 0,
+      cache_creation_input_tokens: cacheWriteTokens,
       service_tier: "standard",
       audio_input_tokens: u.prompt_tokens_details?.audio_tokens ?? 0,
       audio_cache_read_tokens: (u.prompt_tokens_details as any)?.audio_cached_tokens ?? 0,
diff --git a/src/messages/gemini.ts b/src/messages/gemini.ts
index 47b8b92..62d5f6b 100644
--- a/src/messages/gemini.ts
+++ b/src/messages/gemini.ts
@@ -216,6 +216,7 @@ export class MessagesToGeminiConverter {
 
     const promptTokens = promptTokenCount + toolUsePromptTokenCount;
     const completionTokens = candidatesTokenCount + thoughtsTokenCount;
+    const uncachedInputTokens = Math.max(0, promptTokens - cachedContentTokenCount);
 
     return {
       prompt_tokens: promptTokens,
@@ -233,7 +234,7 @@ export class MessagesToGeminiConverter {
       web_search_queries: webSearchQueriesCount,
       tool_use: toolUsePromptTokenCount,
       trafficType,
-      input_tokens: promptTokens,
+      input_tokens: uncachedInputTokens,
       output_tokens: completionTokens,
       cache_read_input_tokens: cachedContentTokenCount,
       cache_creation_input_tokens: 0,
diff --git a/src/messages/responses.ts b/src/messages/responses.ts
index 982f9a2..6ef1b5c 100644
--- a/src/messages/responses.ts
+++ b/src/messages/responses.ts
@@ -170,6 +170,8 @@ export class MessagesToResponsesConverter {
     const outputTokens = u.output_tokens ?? 0;
     const totalTokens = u.total_tokens ?? inputTokens + outputTokens;
     const cachedTokens = u.input_tokens_details?.cached_tokens ?? 0;
+    const cacheWriteTokens = u.input_tokens_details?.cache_write_tokens ?? 0;
+    const uncachedInputTokens = Math.max(0, inputTokens - cachedTokens - cacheWriteTokens);
     const reasoningTokens = u.output_tokens_details?.reasoning_tokens ?? 0;
 
     return {
@@ -185,7 +187,7 @@ export class MessagesToResponsesConverter {
         cached_tokens: cachedTokens,
         ...(webSearchCount > 0 ? { web_search: webSearchCount } : {}),
       },
-      input_tokens: inputTokens,
+      input_tokens: uncachedInputTokens,
       output_tokens: outputTokens,
       cache_read_input_tokens: cachedTokens,
       server_tool_use:
@@ -195,7 +197,7 @@ export class MessagesToResponsesConverter {
               web_search_requests: webSearchCount,
             }
           : null,
-      cache_creation_input_tokens: 0,
+      cache_creation_input_tokens: cacheWriteTokens,
       service_tier: "standard",
       audio_input_tokens: 0,
       audio_cache_read_tokens: 0,