diff --git a/Lib/test/test_opcache.py b/Lib/test/test_opcache.py
index 7946550ec0db63..59f23b86810ce0 100644
--- a/Lib/test/test_opcache.py
+++ b/Lib/test/test_opcache.py
@@ -5,9 +5,11 @@
 import threading
 import types
 import unittest
+import weakref
 from test.support import (threading_helper, check_impl_detail,
-                          requires_specialization,
-                          cpython_only, requires_jit_disabled, reset_code)
+                          requires_specialization, Py_GIL_DISABLED,
+                          cpython_only, requires_jit_disabled, reset_code,
+                          gc_collect)
 from test.support.import_helper import import_module
 
 # Skip this module on other interpreters, it is cpython specific:
@@ -2240,6 +2242,38 @@ def load_enum_member():
         self.assert_specialized(load_enum_member,
                                 "LOAD_ATTR_CLASS_WITH_METACLASS_CHECK")
 
+    @cpython_only
+    @requires_specialization
+    @unittest.skipUnless(Py_GIL_DISABLED,
+                         "deferred refcounting needs the free-threaded build")
+    @threading_helper.requires_working_threading()
+    def test_load_attr_class_mutable_value_deferred_non_owning_threading(self):
+        # Ensure load attributes from class mutable values
+        # are still enabled for deferred referencing
+        # counting from a non owning thread
+        class Value:
+            pass
+
+        class Holder:
+            attr = Value()
+
+        value = Holder.attr
+        wr = weakref.ref(value)
+
+        def load_class_attr():
+            for _ in range(_testinternalcapi.SPECIALIZATION_THRESHOLD):
+                x = Holder.attr
+                assert x is value
+
+        # Ensure we specialise from a non owning thread
+        threading_helper.run_concurrently(load_class_attr, nthreads=1)
+        self.assert_no_opcode(load_class_attr, "LOAD_ATTR_CLASS")
+        self.assertTrue(_testinternalcapi.has_deferred_refcount(value))
+
+        del load_class_attr, Holder, value
+        gc_collect()
+        self.assertIsNone(wr())
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2026-08-15-23-32-52.gh-issue-139103.d8pm7q.rst b/Misc/NEWS.d/next/Core_and_Builtins/2026-08-15-23-32-52.gh-issue-139103.d8pm7q.rst
new file mode 100644
index 00000000000000..0fd62c333c42e3
--- /dev/null
+++ b/Misc/NEWS.d/next/Core_and_Builtins/2026-08-15-23-32-52.gh-issue-139103.d8pm7q.rst
@@ -0,0 +1 @@
+Enable deferred reference counting for mutable class attributes to improve free-threading scaling
diff --git a/Python/specialize.c b/Python/specialize.c
index 773b55c329c772..2695fcb85edc00 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -1216,6 +1216,9 @@ specialize_class_load_attr(PyObject *owner, _Py_CODEUNIT *instr,
             // special case for enums which has Py_TYPE(descr) == cls
             // so guarding on type version is sufficient
             if (Py_TYPE(descr) != cls) {
+#ifdef Py_GIL_DISABLED
+                maybe_enable_deferred_ref_count(descr);
+#endif
                 SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_ATTR_MUTABLE_CLASS);
                 Py_XDECREF(descr);
                 return -1;
diff --git a/Tools/ftscalingbench/ftscalingbench.py b/Tools/ftscalingbench/ftscalingbench.py
index c8a914c22a9e13..609d71ed19df5c 100644
--- a/Tools/ftscalingbench/ftscalingbench.py
+++ b/Tools/ftscalingbench/ftscalingbench.py
@@ -326,6 +326,25 @@ def enum_attr():
         MyEnum.Z
 
 
+class _SharedAttrValue:
+    pass
+
+class MyClassWithSharedAttr:
+    # A class attribute whose value is an instance of a *different* class.
+    # Reading it from multiple threads did not scale because LOAD_ATTR_CLASS
+    # cannot specialize this case (the value's type is not the owner class),
+    # leaving the shared value's reference count contended on every read.
+    attr = _SharedAttrValue()
+
+@register_benchmark
+def class_attribute():
+    obj = MyClassWithSharedAttr
+    for _ in range(1000 * WORK_SCALE):
+        obj.attr
+        obj.attr
+        obj.attr
+
+
 def bench_one_thread(func):
     t0 = time.perf_counter_ns()
     func()