diff --git a/doc/source/release.rst b/doc/source/release.rst
index 5240c67fc42d8..1e1fe4f52a73f 100644
--- a/doc/source/release.rst
+++ b/doc/source/release.rst
@@ -169,6 +169,7 @@ Bug Fixes
 - Bug in :meth:`DataFrame.replace` where nested dicts were erroneously
   depending on the order of dictionary keys and values (:issue:`5338`).
 - Perf issue in concatting with empty objects (:issue:`3259`)
+- Clarify sorting of ``sym_diff`` on ``Index``es with ``NaN``s (:isssue:`6444`)
 
 pandas 0.13.1
 -------------
diff --git a/pandas/core/index.py b/pandas/core/index.py
index 3d821f37e41b5..6c45fccda12ab 100644
--- a/pandas/core/index.py
+++ b/pandas/core/index.py
@@ -1045,6 +1045,9 @@ def sym_diff(self, other, result_name=None):
         ``idx2`` but not both. Equivalent to the Index created by
         ``(idx1 - idx2) + (idx2 - idx1)`` with duplicates dropped.
 
+        The sorting of a result containing ``NaN``s is not guaranteed
+        across Python versions. See GitHub issue #6444.
+
         Examples
         --------
         >>> idx1 = Index([1, 2, 3, 4])
@@ -1067,7 +1070,6 @@ def sym_diff(self, other, result_name=None):
         the_diff = sorted(set((self - other) + (other - self)))
         return Index(the_diff, name=result_name)
 
-
     def unique(self):
         """
         Return array of unique values in the Index. Significantly faster than
diff --git a/pandas/tests/test_index.py b/pandas/tests/test_index.py
index d8625c8687f79..e828bc100dfcf 100644
--- a/pandas/tests/test_index.py
+++ b/pandas/tests/test_index.py
@@ -493,13 +493,16 @@ def test_symmetric_diff(self):
         self.assert_(tm.equalContents(result, expected))
 
         # nans:
-        idx1 = Index([1, 2, np.nan])
+        # GH #6444, sorting of nans. Make sure the number of nans is right
+        # and the correct non-nan values are there. punt on sorting.
+        idx1 = Index([1, 2, 3, np.nan])
         idx2 = Index([0, 1, np.nan])
         result = idx1.sym_diff(idx2)
-        expected = Index([0.0, np.nan, 2.0, np.nan])  # oddness with nans
-        nans = pd.isnull(expected)
-        self.assert_(pd.isnull(result[nans]).all())
-        self.assert_(tm.equalContents(result[~nans], expected[~nans]))
+        # expected = Index([0.0, np.nan, 2.0, 3.0, np.nan])
+        nans = pd.isnull(result)
+        self.assertEqual(nans.sum(), 2)
+        self.assertEqual((~nans).sum(), 3)
+        [self.assertIn(x, result) for x in [0.0, 2.0, 3.0]]
 
         # other not an Index:
         idx1 = Index([1, 2, 3, 4], name='idx1')