diff --git a/mlx/backend/metal/compiled.cpp b/mlx/backend/metal/compiled.cpp
index cda06143d6..e95d7b8d5f 100644
--- a/mlx/backend/metal/compiled.cpp
+++ b/mlx/backend/metal/compiled.cpp
@@ -208,7 +208,7 @@ inline void build_kernel(
         "  {0} tmp_{1} = ", get_type_string(x.dtype()), namer.get_name(x));
     if (is_static_cast(x.primitive())) {
       os += fmt::format(
-          "static_cast<{0}>(tmp_{1});\n",
+          "cast_to<{0}>(tmp_{1});\n",
           get_type_string(x.dtype()),
           namer.get_name(x.inputs()[0]));
     } else {
diff --git a/python/tests/test_compile.py b/python/tests/test_compile.py
index 5eaa6cb955..663ec295b0 100644
--- a/python/tests/test_compile.py
+++ b/python/tests/test_compile.py
@@ -1623,6 +1623,16 @@ def test_compile_abs_unsigned(self):
             x = mx.array([1, 2, 3], dtype)
             self.assertTrue(mx.array_equal(mx.compile(fun)(x), fun(x)))
 
+    def test_compiled_subnormal_bool_cast(self):
+        f32_sub = mx.array(np.array([0x00000001] * 4, dtype=np.uint32)).view(mx.float32)
+        f16_sub = mx.array(np.array([0x0001] * 4, dtype=np.uint16)).view(mx.float16)
+        bf16_sub = mx.array(np.array([0x0001] * 4, dtype=np.uint16)).view(mx.bfloat16)
+
+        # A single-op compile does not fuse; the fused path needs >= 2 ops.
+        fn = mx.compile(lambda x: mx.broadcast_to(x, (2, 4)).astype(mx.bool_))
+        for sub in (f32_sub, f16_sub, bf16_sub):
+            self.assertTrue(mx.all(fn(sub)).item())
+
     def test_compile_different_log_bases(self):
         # The logs are intermediates, since outputs are not simplified.
         def entropies(p):