diff --git a/tests/test_litellm/litellm_core_utils/test_safe_json_loads.py b/tests/test_litellm/litellm_core_utils/test_safe_json_loads.py
new file mode 100644
index 00000000000..5fdc8df15ae
--- /dev/null
+++ b/tests/test_litellm/litellm_core_utils/test_safe_json_loads.py
@@ -0,0 +1,58 @@
+"""
+Unit tests for litellm.litellm_core_utils.safe_json_loads.safe_json_loads.
+
+safe_json_loads parses a JSON string and, on any failure, returns a default
+value instead of raising. These tests cover the success path, the fallback
+path, and the non-string edge case.
+"""
+
+import pytest
+
+from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
+
+
+def test_parses_valid_json_object():
+    assert safe_json_loads('{"a": 1, "b": [2, 3]}') == {"a": 1, "b": [2, 3]}
+
+
+def test_parses_valid_json_array():
+    assert safe_json_loads("[1, 2, 3]") == [1, 2, 3]
+
+
+@pytest.mark.parametrize(
+    "raw, expected",
+    [
+        ('"hello"', "hello"),
+        ("42", 42),
+        ("3.14", 3.14),
+        ("true", True),
+        ("false", False),
+    ],
+)
+def test_parses_valid_json_scalars(raw, expected):
+    assert safe_json_loads(raw) == expected
+
+
+def test_parses_json_null_to_none():
+    # Use a non-None default so a correct parse (real None) is distinguishable
+    # from the function silently falling back to its default.
+    sentinel = object()
+    assert safe_json_loads("null", default=sentinel) is None
+
+
+def test_returns_default_none_on_invalid_json():
+    assert safe_json_loads("{not valid json") is None
+
+
+def test_returns_custom_default_on_invalid_json():
+    sentinel = {"fallback": True}
+    assert safe_json_loads("oops", default=sentinel) is sentinel
+
+
+def test_returns_default_on_empty_string():
+    assert safe_json_loads("") is None
+
+
+@pytest.mark.parametrize("bad_input", [None, 123, 4.5, ["a"], {"b": 1}])
+def test_returns_default_on_non_string_input(bad_input):
+    assert safe_json_loads(bad_input, default="fallback") == "fallback"