diff --git a/changelog.d/fixes/10591-gemini-live-catalog.md b/changelog.d/fixes/10591-gemini-live-catalog.md
new file mode 100644
index 0000000000..7663591649
--- /dev/null
+++ b/changelog.d/fixes/10591-gemini-live-catalog.md
@@ -0,0 +1 @@
+- Stop advertising Gemini Live-only models as supported audio endpoints until OmniRoute proxies the bidirectional Live protocol.
diff --git a/src/lib/providerModels/geminiModelsParser.ts b/src/lib/providerModels/geminiModelsParser.ts
index e4fd1bfdd7..9d4aa68de8 100644
--- a/src/lib/providerModels/geminiModelsParser.ts
+++ b/src/lib/providerModels/geminiModelsParser.ts
@@ -6,7 +6,7 @@
  *   - predict                          → "images"  (Imagen image generation)
  *   - predictLongRunning               → "videos"  (Veo video generation)
  *   - embedContent                     → "embeddings"
- *   - bidiGenerateContent              → "audio"   (Live real-time audio)
+ *   - bidiGenerateContent              → ignored   (Gemini Live is not proxied)
  *
  * Model-id heuristics refine the long-running bucket because Google exposes both
  * Imagen and Veo via long-running methods on the same endpoint:
@@ -19,7 +19,7 @@
  *
  * This is shared by the `gemini` discovery config and the `vertex` /
  * `vertex-partner` (incl. Vertex AI Express key) discovery branches, so every
- * model the account can access — chat, image, video, audio and embeddings —
+ * supported model the account can access — chat, image, video and embeddings —
  * surfaces dynamically instead of being limited to the small static registry.
  */
 const METHOD_TO_ENDPOINT: Record<string, string> = {
@@ -27,7 +27,6 @@ const METHOD_TO_ENDPOINT: Record<string, string> = {
   embedContent: "embeddings",
   predict: "images",
   predictLongRunning: "videos",
-  bidiGenerateContent: "audio",
   generateAnswer: "chat",
 };
 
@@ -37,6 +36,7 @@ const IGNORED_METHODS = new Set([
   "createCachedContent",
   "batchGenerateContent",
   "asyncBatchEmbedContent",
+  "bidiGenerateContent",
 ]);
 
 const RETIRED_GEMINI_MODEL_IDS = new Set(["gemini-3.5-flash"]);
@@ -78,6 +78,13 @@ export function parseGeminiModelsList(data: any): GeminiDiscoveryModel[] {
         endpoints.add("images");
       }
 
+      if (
+        endpoints.size === 0 &&
+        methods.length > 0 &&
+        methods.every((method) => IGNORED_METHODS.has(method))
+      ) {
+        return null;
+      }
       if (endpoints.size === 0) endpoints.add("chat");
 
       return {
@@ -91,5 +98,8 @@ export function parseGeminiModelsList(data: any): GeminiDiscoveryModel[] {
         ...(m.thinking === true ? { supportsThinking: true } : {}),
       } as GeminiDiscoveryModel;
     })
-    .filter((model: GeminiDiscoveryModel) => !RETIRED_GEMINI_MODEL_IDS.has(model.id));
+    .filter(
+      (model: GeminiDiscoveryModel | null): model is GeminiDiscoveryModel =>
+        Boolean(model) && !RETIRED_GEMINI_MODEL_IDS.has(model.id)
+    );
 }
diff --git a/tests/unit/gemini-models-parser.test.ts b/tests/unit/gemini-models-parser.test.ts
index 8c9da3dac4..f9463d61b5 100644
--- a/tests/unit/gemini-models-parser.test.ts
+++ b/tests/unit/gemini-models-parser.test.ts
@@ -72,14 +72,21 @@ test("parseGeminiModelsList maps generateContent image models to the chat endpoi
   assert.deepEqual(proImage!.supportedEndpoints, ["chat"]);
 });
 
-test("parseGeminiModelsList maps embedContent and bidiGenerateContent", () => {
+test("parseGeminiModelsList maps embeddings without advertising unsupported Gemini Live", () => {
   const models = parseGeminiModelsList(SAMPLE);
   assert.deepEqual(models.find((m) => m.id === "text-embedding-004")!.supportedEndpoints, [
     "embeddings",
   ]);
-  assert.deepEqual(models.find((m) => m.id === "gemini-live-2.5-flash")!.supportedEndpoints, [
-    "audio",
-  ]);
+  assert.equal(models.some((m) => m.id === "gemini-live-2.5-flash"), false);
+  const [hybrid] = parseGeminiModelsList({
+    models: [
+      {
+        name: "models/gemini-live-hybrid",
+        supportedGenerationMethods: ["generateContent", "bidiGenerateContent"],
+      },
+    ],
+  });
+  assert.deepEqual(hybrid.supportedEndpoints, ["chat"]);
 });
 
 test("parseGeminiModelsList maps Veo predictLongRunning models to the videos endpoint", () => {