diff --git a/gateway/platforms/base.py b/gateway/platforms/base.py
index 417893fea2d4b..c8fad2c9f93f6 100644
--- a/gateway/platforms/base.py
+++ b/gateway/platforms/base.py
@@ -3118,6 +3118,14 @@ def truncate_message(
                     if safe_split > _cp_limit // 4:
                         split_at = safe_split
 
+            if _len is not len:
+                # Re-clamp after natural-break adjustments so the final slice
+                # is still valid under custom unit counting (e.g. UTF-16).
+                safe_prefix = _prefix_within_utf16_limit(remaining[:split_at], headroom)
+                split_at = len(safe_prefix)
+                if split_at < 1:
+                    split_at = max(1, _custom_unit_to_cp(remaining, max(1, headroom), _len))
+
             chunk_body = remaining[:split_at]
             remaining = remaining[split_at:].lstrip()
 
diff --git a/tests/gateway/test_platform_base.py b/tests/gateway/test_platform_base.py
index a6e0d51d60eeb..14a3bb2e11cca 100644
--- a/tests/gateway/test_platform_base.py
+++ b/tests/gateway/test_platform_base.py
@@ -633,6 +633,18 @@ def test_code_blocks_preserved_with_utf16(self):
                 f"Chunk {i} has unbalanced fences ({fence_count})"
             )
 
+    def test_natural_break_split_keeps_utf16_safe_boundaries(self):
+        """Natural break preference must still produce UTF-16-safe chunks."""
+        msg = ("alpha 🎩 beta " * 400).strip()
+        chunks = BasePlatformAdapter.truncate_message(msg, 120, len_fn=utf16_len)
+        assert len(chunks) > 1
+        for i, chunk in enumerate(chunks):
+            # Regression guard: chunk should always be encodable and within limits.
+            chunk.encode("utf-16-le")
+            assert utf16_len(chunk) <= 120 + 20, (
+                f"Chunk {i} UTF-16 length {utf16_len(chunk)} exceeds limit"
+            )
+
 
 class TestProxyKwargsForAiohttp:
     """Verify proxy_kwargs_for_aiohttp routes all schemes through ProxyConnector."""