diff --git a/Include/internal/pycore_long.h b/Include/internal/pycore_long.h
index fb5622c99f7a135..bbaf59aab46d638 100644
--- a/Include/internal/pycore_long.h
+++ b/Include/internal/pycore_long.h
@@ -117,6 +117,12 @@ PyAPI_FUNC(_PyStackRef) _PyCompactLong_Add(PyLongObject *left, PyLongObject *rig
 PyAPI_FUNC(_PyStackRef) _PyCompactLong_Multiply(PyLongObject *left, PyLongObject *right);
 PyAPI_FUNC(_PyStackRef) _PyCompactLong_Subtract(PyLongObject *left, PyLongObject *right);
 
+// Add or subtract two int objects, calling the long implementation directly
+// instead of going through the number protocol. Used by the BINARY_OP_EXTEND
+// specialization for operands that are not both compact.
+extern PyObject* _PyLong_Add(PyObject *left, PyObject *right);
+extern PyObject* _PyLong_Subtract(PyObject *left, PyObject *right);
+
 // Export for 'binascii' shared extension.
 PyAPI_DATA(unsigned char) _PyLong_DigitValue[256];
 
diff --git a/Lib/test/test_capi/test_opt.py b/Lib/test/test_capi/test_opt.py
index 88077aacecbf5b1..8ea34db410ebdbc 100644
--- a/Lib/test/test_capi/test_opt.py
+++ b/Lib/test/test_capi/test_opt.py
@@ -4887,6 +4887,47 @@ def testfunc(n):
         self.assertIn("_UNPACK_SEQUENCE_TUPLE", uops)
         self.assertNotIn("_GUARD_TOS_TUPLE", uops)
 
+    def test_binary_op_extend_wide_int_subtract(self):
+        # Non-compact int - int is specialized via BINARY_OP_EXTEND. The
+        # result is known to be an int, so the second subtraction only needs
+        # to guard its rhs.
+        def testfunc(n):
+            a = 10**20
+            b = 10**19
+            c = 10**18
+            res = 0
+            for _ in range(n):
+                res = (a - b) - c
+            return res
+
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, 10**20 - 10**19 - 10**18)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertEqual(uops.count("_BINARY_OP_EXTEND"), 2)
+        self.assertEqual(uops.count("_GUARD_BINARY_OP_EXTEND"), 1)
+        self.assertIn("_GUARD_BINARY_OP_EXTEND_RHS", uops)
+        self.assertNotIn("_BINARY_OP_SUBTRACT_INT", uops)
+
+    def test_compact_int_ops_see_non_compact_operands(self):
+        # Trace compiled for compact ints, then re-entered with operands whose
+        # difference is not compact. The follow-up int op must not treat the
+        # result as compact.
+        def testfunc(args):
+            a, b, c, n = args
+            res = 0
+            for _ in range(n):
+                res = (a - b) - c
+            return res
+
+        res, ex = self._run_with_optimizer(testfunc, (5, 2, 1000, TIER2_THRESHOLD))
+        self.assertEqual(res, -997)
+        self.assertIsNotNone(ex)
+        self.assertIn("_BINARY_OP_SUBTRACT_INT", get_opnames(ex))
+        a, b = 2**60 + 2**59, 2**60
+        res = testfunc((a, b, 1000, TIER2_THRESHOLD))
+        self.assertEqual(res, 2**59 - 1000)
+
     def test_binary_op_extend_guard_elimination(self):
         # When both operands have known types (e.g., from a prior
         # _BINARY_OP_EXTEND result), the _GUARD_BINARY_OP_EXTEND
diff --git a/Lib/test/test_opcache.py b/Lib/test/test_opcache.py
index 60879e2774e7077..4a42a72080b9813 100644
--- a/Lib/test/test_opcache.py
+++ b/Lib/test/test_opcache.py
@@ -1536,6 +1536,39 @@ def binary_op_bitwise_extend():
         self.assert_specialized(binary_op_bitwise_extend, "BINARY_OP_EXTEND")
         self.assert_no_opcode(binary_op_bitwise_extend, "BINARY_OP")
 
+        def binary_op_wide_int_extend():
+            for _ in range(100):
+                a, b = 10**20, 10**19
+                x = a + b
+                self.assertEqual(x, 11 * 10**19)
+                y = a - b
+                self.assertEqual(y, 9 * 10**19)
+                a += b
+                self.assertEqual(a, 11 * 10**19)
+                a -= b
+                self.assertEqual(a, 10**20)
+
+        binary_op_wide_int_extend()
+        self.assert_specialized(binary_op_wide_int_extend, "BINARY_OP_EXTEND")
+        self.assert_no_opcode(binary_op_wide_int_extend, "BINARY_OP")
+        self.assert_no_opcode(binary_op_wide_int_extend, "BINARY_OP_ADD_INT")
+        self.assert_no_opcode(binary_op_wide_int_extend, "BINARY_OP_SUBTRACT_INT")
+
+        def wide_int_extend_mixed(a, b):
+            return a - b, a + b
+
+        # Specialized on wide operands; compact operands and bool still
+        # produce correct results (bool is not an exact int and deopts).
+        for _ in range(100):
+            self.assertEqual(wide_int_extend_mixed(10**20, 1),
+                             (10**20 - 1, 10**20 + 1))
+        self.assert_specialized(wide_int_extend_mixed, "BINARY_OP_EXTEND")
+        self.assertEqual(wide_int_extend_mixed(5, 7), (-2, 12))
+        self.assertEqual(wide_int_extend_mixed(True, 10**20),
+                         (1 - 10**20, 1 + 10**20))
+        self.assertEqual(wide_int_extend_mixed(-10**20, 10**20),
+                         (-2 * 10**20, 0))
+
     @cpython_only
     @requires_specialization
     def test_load_super_attr(self):
diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2026-10-18-12-00-00.gh-issue-151289.wIdEx7.rst b/Misc/NEWS.d/next/Core_and_Builtins/2026-10-18-12-00-00.gh-issue-151289.wIdEx7.rst
new file mode 100644
index 000000000000000..150bcb709adda19
--- /dev/null
+++ b/Misc/NEWS.d/next/Core_and_Builtins/2026-10-18-12-00-00.gh-issue-151289.wIdEx7.rst
@@ -0,0 +1,3 @@
+Specialize :class:`int` addition and subtraction where an operand is not a
+compact integer, using ``BINARY_OP_EXTEND``. This skips the generic binary
+operator dispatch for these cases.
diff --git a/Objects/longobject.c b/Objects/longobject.c
index 9577c8ef4d6ed08..e71242e2459fc5d 100644
--- a/Objects/longobject.c
+++ b/Objects/longobject.c
@@ -3923,6 +3923,20 @@ long_sub_method(PyObject *a, PyObject *b)
     return (PyObject*)long_sub((PyLongObject*)a, (PyLongObject*)b);
 }
 
+PyObject *
+_PyLong_Add(PyObject *a, PyObject *b)
+{
+    assert(PyLong_Check(a) && PyLong_Check(b));
+    return (PyObject *)long_add((PyLongObject *)a, (PyLongObject *)b);
+}
+
+PyObject *
+_PyLong_Subtract(PyObject *a, PyObject *b)
+{
+    assert(PyLong_Check(a) && PyLong_Check(b));
+    return (PyObject *)long_sub((PyLongObject *)a, (PyLongObject *)b);
+}
+
 
 /* Grade school multiplication, ignoring the signs.
  * Returns the absolute value of the product, or NULL if error.
diff --git a/Python/specialize.c b/Python/specialize.c
index f658df3bc5d2228..e5813b42246d833 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -2256,6 +2256,16 @@ static _PyBinaryOpSpecializationDescr binaryop_extend_descrs[] = {
     {NB_INPLACE_AND, compactlongs_guard, compactlongs_and, &PyLong_Type, 1, NULL, NULL},
     {NB_INPLACE_XOR, compactlongs_guard, compactlongs_xor, &PyLong_Type, 1, NULL, NULL},
 
+    /* long-long add/subtract for exact ints that are not both compact (those
+       take BINARY_OP_ADD_INT and friends). Calling the long implementation
+       directly skips the generic binary_op1 slot dispatch. The guard is a
+       pure type check: the result may be non-compact, so it is typed as int
+       without compactness and follow-up int guards stay. */
+    {NB_ADD, NULL, _PyLong_Add, &PyLong_Type, 1, &PyLong_Type, &PyLong_Type},
+    {NB_SUBTRACT, NULL, _PyLong_Subtract, &PyLong_Type, 1, &PyLong_Type, &PyLong_Type},
+    {NB_INPLACE_ADD, NULL, _PyLong_Add, &PyLong_Type, 1, &PyLong_Type, &PyLong_Type},
+    {NB_INPLACE_SUBTRACT, NULL, _PyLong_Subtract, &PyLong_Type, 1, &PyLong_Type, &PyLong_Type},
+
     /* float-long arithmetic: guards also check NaN and compactness. */
     {NB_ADD, float_compactlong_guard, float_compactlong_add, &PyFloat_Type, 1, NULL, NULL},
     {NB_SUBTRACT, float_compactlong_guard, float_compactlong_subtract, &PyFloat_Type, 1, NULL, NULL},