diff --git a/Lib/test/test_curses.py b/Lib/test/test_curses.py
index e15817ce2e7b0f0..a838ea6e123ffce 100644
--- a/Lib/test/test_curses.py
+++ b/Lib/test/test_curses.py
@@ -326,13 +326,10 @@ def test_output_character(self):
                 stdscr.move(2, 0)
                 stdscr.echochar(v)
                 self.assertEqual(self._read_char(2, 0), c)
-                # insch() round-trips a byte only where its code point equals
-                # the byte value (Latin-1): on a wide build ncurses winsch
-                # stores a printable byte directly as a code point instead of
-                # decoding it through the locale.
-                if ord(c) < 0x100:
-                    stdscr.insch(1, 0, v)
-                    self.assertEqual(self._read_char(1, 0), c)
+                # insch() decodes the byte through the locale like addch(), so
+                # it round-trips the same character.
+                stdscr.insch(1, 0, v)
+                self.assertEqual(self._read_char(1, 0), c)
 
         # The same characters supplied as a str.  Unlike the int path above, a
         # str is stored as a wide-character cell on a wide build, so every
diff --git a/Misc/NEWS.d/next/Library/2026-07-17-20-56-32.gh-issue-153864.WmA9By.rst b/Misc/NEWS.d/next/Library/2026-07-17-20-56-32.gh-issue-153864.WmA9By.rst
new file mode 100644
index 000000000000000..79857185d7d0c76
--- /dev/null
+++ b/Misc/NEWS.d/next/Library/2026-07-17-20-56-32.gh-issue-153864.WmA9By.rst
@@ -0,0 +1,3 @@
+On a wide :mod:`curses` build, :meth:`curses.window.insch` now inserts a
+non-ASCII byte as the character it encodes in the window's encoding,
+consistently with :meth:`~curses.window.addch`, instead of its code point.
diff --git a/Modules/_cursesmodule.c b/Modules/_cursesmodule.c
index 51c4854aa5da52d..2b36ed7f96fc783 100644
--- a/Modules/_cursesmodule.c
+++ b/Modules/_cursesmodule.c
@@ -1871,6 +1871,27 @@ _curses_window_insch_impl(PyCursesWindowObject *self, int group_left_1,
     if (!PyCurses_ConvertToChtype(self, ch, &ch_))
         return NULL;
 
+#ifdef HAVE_NCURSESW
+    /* winsch() does not locale-decode a byte above 127 on a wide build,
+       unlike waddch(), so decode it here and insert it as a wide character. */
+    chtype cch = ch_ & A_CHARTEXT;
+    if (cch > 127) {
+        wint_t wc = btowc((int)cch);
+        if (wc != WEOF) {
+            cchar_t wch;
+            wchar_t wstr[2] = { (wchar_t)wc, L'\0' };
+            attr_t cattr = (attr_t)((ch_ | (attr_t)attr) & ~(chtype)A_CHARTEXT);
+            setcchar(&wch, wstr, cattr, PAIR_NUMBER(cattr), NULL);
+            if (!group_left_1) {
+                rtn = wins_wch(self->win, &wch);
+            }
+            else {
+                rtn = mvwins_wch(self->win, y, x, &wch);
+            }
+            return PyCursesCheckERR_ForWin(self, rtn, "insch");
+        }
+    }
+#endif
     if (!group_left_1) {
         rtn = winsch(self->win, ch_ | (attr_t)attr);
     }