diff --git a/tests/e2e/llm_translation/fixtures/cat.jpg b/tests/e2e/llm_translation/fixtures/cat.jpg
new file mode 100644
index 00000000000..103c370b2e2
Binary files /dev/null and b/tests/e2e/llm_translation/fixtures/cat.jpg differ
diff --git a/tests/e2e/llm_translation/realtime/REALTIME_COVERAGE_MATRIX.md b/tests/e2e/llm_translation/realtime/REALTIME_COVERAGE_MATRIX.md
index bae858d50af..a6e32b88479 100644
--- a/tests/e2e/llm_translation/realtime/REALTIME_COVERAGE_MATRIX.md
+++ b/tests/e2e/llm_translation/realtime/REALTIME_COVERAGE_MATRIX.md
@@ -42,7 +42,7 @@ at call time. The provider table below is the source of truth; edit `PROVIDERS`
 | openai | `openai-realtime` | `openai/gpt-realtime-2` |
 | azure | `azure-realtime` | `azure/gpt-realtime-2` (GA protocol) |
 | gemini | `gemini-realtime` | `gemini/gemini-3.1-flash-live-preview` |
-| vertex_ai | `vertex-realtime` | `vertex_ai/gemini-live-2.5-flash-preview-native-audio-09-2025` |
+| vertex_ai | `vertex-realtime` | `vertex_ai/gemini-live-2.5-flash-native-audio` |
 
 Bedrock and xai (`xai/grok-4-1-fast-non-reasoning`) are supported by the proxy but
 kept commented out in `PROVIDERS` until they pass end-to-end here; re-enable them by
diff --git a/tests/e2e/llm_translation/realtime/realtime_client.py b/tests/e2e/llm_translation/realtime/realtime_client.py
index 632a9cf7e57..3ffca7e8b88 100644
--- a/tests/e2e/llm_translation/realtime/realtime_client.py
+++ b/tests/e2e/llm_translation/realtime/realtime_client.py
@@ -78,7 +78,7 @@ PROVIDERS = (
         "vertex_ai",
         "vertex-realtime",
         LiteLLMParamsBody(
-            model="vertex_ai/gemini-live-2.5-flash-preview-native-audio-09-2025",
+            model="vertex_ai/gemini-live-2.5-flash-native-audio",
             vertex_location="us-central1",
             vertex_credentials="os.environ/VERTEXAI_CREDENTIALS",
         ),
diff --git a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py
index 655d426c28d..fe5cfd8f2ec 100644
--- a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py
+++ b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py
@@ -16,7 +16,10 @@ via /model/new (Cohere, Gemini, hosted_vllm), each deleted on teardown.
 
 from __future__ import annotations
 
+import base64
 import os
+from pathlib import Path
+from typing import Final
 
 import pytest
 from pydantic import BaseModel
@@ -79,18 +82,29 @@ def _streamed_tool_call(events: list[str]) -> tuple[str, str]:
     return name, arguments
 
 
-CAT_IMAGE_URL = "https://upload.wikimedia.org/wikipedia/commons/3/3a/Cat03.jpg"
+_FIXTURES_DIR: Final = Path(__file__).parent / "fixtures"
+CAT_IMAGE: Final = _FIXTURES_DIR / "cat.jpg"
 OPENAI_VISION_BACKEND = "openai/gpt-4o"
 
-# OpenAI caches a shared prompt prefix once it exceeds ~1024 tokens; this is well
-# past that, so a repeat call reports cached prompt tokens.
+
+def _cat_image_data_url() -> str:
+    """The vision image as a data URL, read from a fixture we own.
+
+    An https URL would make every vision run depend on a third-party host staying
+    up and unthrottled, and a 429 from that host reads as a gateway failure. It also
+    changes what is under test per provider: litellm downloads the image itself for
+    bedrock, while openai is handed the link and fetches it from its own servers. A
+    data URL removes the host and puts both providers on the same bytes."""
+    return "data:image/jpeg;base64," + base64.b64encode(CAT_IMAGE.read_bytes()).decode()
+
+
 def _vision_messages() -> list[ChatMessage]:
     return [
         ChatMessage(
             role="user",
             content=[
                 TextContentPart(text="What animal is in this image? Answer in one word."),
-                ImageContentPart(image_url=ImageUrl(url=CAT_IMAGE_URL)),
+                ImageContentPart(image_url=ImageUrl(url=_cat_image_data_url())),
             ],
         )
     ]