diff --git a/Misc/NEWS b/Misc/NEWS
index 3dd58e6..472751f 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -12,6 +12,11 @@
 Core and Builtins
 -----------------
 
+- Issue #4445: Replace "sizeof(PyBytesObject)" with
+  "offsetof(PyBytesObject, ob_sval) + 1" when allocating memory for
+  bytes instances.  On a typical machine this saves 3 bytes of memory
+  (on average) per allocation of a bytes instance.
+
 - Issue #4533: File read operation was dreadfully slow due to a slowly
   growing read buffer. Fixed by using the same growth rate algorithm as
   Python 2.x.
diff --git a/Objects/bytesobject.c b/Objects/bytesobject.c
index 76b7f52..ca8b857 100644
--- a/Objects/bytesobject.c
+++ b/Objects/bytesobject.c
@@ -5,6 +5,7 @@
 #include "Python.h"
 
 #include "bytes_methods.h"
+#include <stddef.h>
 
 static Py_ssize_t
 _getbuffer(PyObject *obj, Py_buffer *view)
@@ -31,6 +32,14 @@
 static PyBytesObject *characters[UCHAR_MAX + 1];
 static PyBytesObject *nullstring;
 
+/* PyBytesObject_SIZE gives the basic size of a string; any memory allocation
+   for a string of length n should request PyBytesObject_SIZE + n bytes.
+
+   Using PyBytesObject_SIZE instead of sizeof(PyBytesObject) saves
+   3 bytes per string allocation on a typical system.
+*/
+#define PyBytesObject_SIZE (offsetof(PyBytesObject, ob_sval) + 1)
+
 /*
    For both PyBytes_FromString() and PyBytes_FromStringAndSize(), the
    parameter `size' denotes number of characters to allocate, not counting any
@@ -83,14 +92,14 @@
 		return (PyObject *)op;
 	}
 
-	if (size > PY_SSIZE_T_MAX - sizeof(PyBytesObject)) {
+	if (size > PY_SSIZE_T_MAX - PyBytesObject_SIZE) {
 		PyErr_SetString(PyExc_OverflowError,
 				"byte string is too large");
 		return NULL;
 	}
 
 	/* Inline PyObject_NewVar */
-	op = (PyBytesObject *)PyObject_MALLOC(sizeof(PyBytesObject) + size);
+	op = (PyBytesObject *)PyObject_MALLOC(PyBytesObject_SIZE + size);
 	if (op == NULL)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyBytes_Type, size);
@@ -117,7 +126,7 @@
 
 	assert(str != NULL);
 	size = strlen(str);
-	if (size > PY_SSIZE_T_MAX - sizeof(PyBytesObject)) {
+	if (size > PY_SSIZE_T_MAX - PyBytesObject_SIZE) {
 		PyErr_SetString(PyExc_OverflowError,
 			"byte string is too long");
 		return NULL;
@@ -138,7 +147,7 @@
 	}
 
 	/* Inline PyObject_NewVar */
-	op = (PyBytesObject *)PyObject_MALLOC(sizeof(PyBytesObject) + size);
+	op = (PyBytesObject *)PyObject_MALLOC(PyBytesObject_SIZE + size);
 	if (op == NULL)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyBytes_Type, size);
@@ -746,13 +755,12 @@
 		return (PyObject *)a;
 	}
 	nbytes = (size_t)size;
-	if (nbytes + sizeof(PyBytesObject) <= nbytes) {
+	if (nbytes + PyBytesObject_SIZE <= nbytes) {
 		PyErr_SetString(PyExc_OverflowError,
 			"repeated bytes are too long");
 		return NULL;
 	}
-	op = (PyBytesObject *)
-		PyObject_MALLOC(sizeof(PyBytesObject) + nbytes);
+	op = (PyBytesObject *)PyObject_MALLOC(PyBytesObject_SIZE + nbytes);
 	if (op == NULL)
 		return PyErr_NoMemory();
 	PyObject_INIT_VAR(op, &PyBytes_Type, size);
@@ -2803,7 +2811,7 @@
 string_sizeof(PyBytesObject *v)
 {
 	Py_ssize_t res;
-	res = sizeof(PyBytesObject) + Py_SIZE(v) * Py_TYPE(v)->tp_itemsize;
+	res = PyBytesObject_SIZE + Py_SIZE(v) * Py_TYPE(v)->tp_itemsize;
 	return PyLong_FromSsize_t(res);
 }
 
@@ -3080,7 +3088,7 @@
 PyTypeObject PyBytes_Type = {
 	PyVarObject_HEAD_INIT(&PyType_Type, 0)
 	"bytes",
-	sizeof(PyBytesObject),
+	PyBytesObject_SIZE,
 	sizeof(char),
  	string_dealloc, 			/* tp_dealloc */
 	0,			 		/* tp_print */
@@ -3175,7 +3183,7 @@
 	_Py_DEC_REFTOTAL;
 	_Py_ForgetReference(v);
 	*pv = (PyObject *)
-		PyObject_REALLOC((char *)v, sizeof(PyBytesObject) + newsize);
+		PyObject_REALLOC((char *)v, PyBytesObject_SIZE + newsize);
 	if (*pv == NULL) {
 		PyObject_Del(v);
 		PyErr_NoMemory();