diff --git a/src/_pytest/python_api.py b/src/_pytest/python_api.py
index f35351e2a..04022b9fe 100644
--- a/src/_pytest/python_api.py
+++ b/src/_pytest/python_api.py
@@ -211,10 +211,9 @@ class ApproxScalar(ApproxBase):
         the pre-specified tolerance.
         """
         if _is_numpy_array(actual):
-            return (
-                ApproxNumpy(actual, rel=self.rel, abs=self.abs, nan_ok=self.nan_ok)
-                == self.expected
-            )
+            import numpy as np
+
+            return np.all(abs(self.expected - actual) <= self.tolerance)
 
         # Short-circuit exact equality.
         if actual == self.expected:
diff --git a/testing/python/approx.py b/testing/python/approx.py
index b93ff241b..09487cc73 100644
--- a/testing/python/approx.py
+++ b/testing/python/approx.py
@@ -346,15 +346,22 @@ class TestApprox(object):
         """
         quick check that numpy rel/abs args are handled correctly
         for comparison against an np.array
+        - 3.6.4 would approx both actual / expected if np.array
+         regardless of which value was passed to approx()
+         Means tolerance could be calculated against bad test return
         """
         np = pytest.importorskip("numpy")
         expected = 100
         actual = 99
         assert actual != pytest.approx(expected, abs=0.1, rel=0)
         assert np.array(actual) != pytest.approx(expected, abs=0.1, rel=0)
+        assert actual != pytest.approx(np.array(expected), abs=0.1, rel=0)
+        assert np.array(actual) != pytest.approx(np.array(expected), abs=0.1, rel=0)
 
         assert actual == pytest.approx(expected, abs=0, rel=0.01)
-        assert np.array(actual) == pytest.approx(expected, abs=0, rel=0.1)
+        assert np.array(actual) == pytest.approx(expected, abs=0, rel=0.01)
+        assert actual == pytest.approx(np.array(expected), abs=0, rel=0.01)
+        assert np.array(actual) == pytest.approx(np.array(expected), abs=0, rel=0.01)
 
     def test_numpy_array_wrong_shape(self):
         np = pytest.importorskip("numpy")