From 7a18f344b368e9da72b4587ada1aede8a5f9c630 Mon Sep 17 00:00:00 2001
From: Thiago Mafra <mfryh1537@proton.me>
Date: Fri, 18 Sep 2026 11:29:38 -0300
Subject: [PATCH] fix(translator): stop double-counting cached tokens in Gemini
 to Claude usage (#12863)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* fix(translator): stop double-counting cached tokens in Gemini→Claude usage

* docs(changelog): add fragment for gemini-to-claude cached input tokens fix
---
 ...63-gemini-to-claude-cached-input-tokens.md |  1 +
 .../translator/response/gemini-to-claude.ts   |  2 +-
 .../translator-resp-gemini-to-claude.test.ts  | 51 ++++++++++++++++++-
 3 files changed, 52 insertions(+), 2 deletions(-)
 create mode 100644 changelog.d/fixes/12863-gemini-to-claude-cached-input-tokens.md

diff --git a/changelog.d/fixes/12863-gemini-to-claude-cached-input-tokens.md b/changelog.d/fixes/12863-gemini-to-claude-cached-input-tokens.md
new file mode 100644
index 0000000000..915d5ffa68
--- /dev/null
+++ b/changelog.d/fixes/12863-gemini-to-claude-cached-input-tokens.md
@@ -0,0 +1 @@
+- **fix(translator):** Gemini to Claude usage no longer double-counts the cached prompt prefix — `input_tokens` now excludes `cache_read_input_tokens`, matching the Anthropic Messages semantics ([#12863](https://github.com/diegosouzapw/OmniRoute/pull/12863))
diff --git a/open-sse/translator/response/gemini-to-claude.ts b/open-sse/translator/response/gemini-to-claude.ts
index 8b06d8707a..523f55d6b1 100644
--- a/open-sse/translator/response/gemini-to-claude.ts
+++ b/open-sse/translator/response/gemini-to-claude.ts
@@ -397,7 +397,7 @@ export function geminiToClaudeResponse(chunk, state) {
       typeof usageMeta.cachedContentTokenCount === "number" ? usageMeta.cachedContentTokenCount : 0;
 
     state.usage = {
-      input_tokens: inputTokens,
+      input_tokens: Math.max(0, inputTokens - cachedTokens),
       output_tokens: candidatesTokens + thoughtsTokens,
     };
     if (cachedTokens > 0) {
diff --git a/tests/unit/translator-resp-gemini-to-claude.test.ts b/tests/unit/translator-resp-gemini-to-claude.test.ts
index 4e63824842..85917db19b 100644
--- a/tests/unit/translator-resp-gemini-to-claude.test.ts
+++ b/tests/unit/translator-resp-gemini-to-claude.test.ts
@@ -107,7 +107,7 @@ test("Gemini -> Claude stream: functionCall becomes tool_use and MAX_TOKENS maps
   assert.equal(result[2].delta.partial_json, JSON.stringify({ path: "/tmp/a" }));
   assert.equal(result[3].type, "content_block_stop");
   assert.equal(result[4].delta.stop_reason, "tool_use");
-  assert.equal(result[4].usage.input_tokens, 5);
+  assert.equal(result[4].usage.input_tokens, 4);
   assert.equal(result[4].usage.output_tokens, 5);
   assert.equal(result[4].usage.cache_read_input_tokens, 1);
   assert.equal(result[5].type, "message_stop");
@@ -190,3 +190,52 @@ test("Gemini -> Claude stream: response wrapper is supported and promptFeedback-
   assert.equal(wrapped[2].delta.text, "wrapped");
   assert.equal(geminiToClaudeResponse({ promptFeedback: { blockReason: "SAFETY" } }, {}), null);
 });
+
+test("Gemini -> Claude stream: input_tokens excludes cached tokens", () => {
+  const result = geminiToClaudeResponse(
+    {
+      responseId: "resp-6",
+      modelVersion: "gemini-2.5-pro",
+      candidates: [
+        {
+          content: { parts: [{ text: "done" }] },
+          finishReason: "STOP",
+        },
+      ],
+      usageMetadata: {
+        promptTokenCount: 100,
+        candidatesTokenCount: 7,
+        cachedContentTokenCount: 90,
+      },
+    },
+    {}
+  );
+
+  const delta = result.find((event) => event.type === "message_delta");
+  assert.equal(delta.usage.input_tokens, 10);
+  assert.equal(delta.usage.cache_read_input_tokens, 90);
+});
+
+test("Gemini -> Claude stream: input_tokens equals promptTokenCount when nothing is cached", () => {
+  const result = geminiToClaudeResponse(
+    {
+      responseId: "resp-7",
+      modelVersion: "gemini-2.5-pro",
+      candidates: [
+        {
+          content: { parts: [{ text: "done" }] },
+          finishReason: "STOP",
+        },
+      ],
+      usageMetadata: {
+        promptTokenCount: 100,
+        candidatesTokenCount: 7,
+      },
+    },
+    {}
+  );
+
+  const delta = result.find((event) => event.type === "message_delta");
+  assert.equal(delta.usage.input_tokens, 100);
+  assert.equal(delta.usage.cache_read_input_tokens, undefined);
+});