diff --git a/litellm/router_utils/fallback_event_handlers.py b/litellm/router_utils/fallback_event_handlers.py
index 4c8688a2d0b..fb390462728 100644
--- a/litellm/router_utils/fallback_event_handlers.py
+++ b/litellm/router_utils/fallback_event_handlers.py
@@ -48,6 +48,12 @@ def _resolve_dict_fallback_targets(
     directly rather than calling that helper because the helper pops string
     entries mid-iteration, which skips the following entry and can drop a dict
     target when a bare string precedes it in the list.
+
+    Within each tier this matches the helper's pick: exact match breaks on the
+    first hit (first wins), while stripped and wildcard hits keep overwriting
+    without a break (last wins). Mirroring last-wins for those tiers keeps the
+    capability scan aligned with the target the runtime fallback hop will
+    actually call.
     """
     dict_entries = tuple(
         (key, entry[key])
@@ -58,18 +64,17 @@ def _resolve_dict_fallback_targets(
     exact = next((t for k, t in dict_entries if k == model_group), None)
     if exact is not None:
         return tuple(exact)
-    stripped = next(
-        (
-            t
-            for k, t in dict_entries
-            if _check_stripped_model_group(model_group=model_group, fallback_key=k)
-        ),
-        None,
+    stripped_matches = tuple(
+        t
+        for k, t in dict_entries
+        if _check_stripped_model_group(model_group=model_group, fallback_key=k)
     )
-    if stripped is not None:
-        return tuple(stripped)
-    wildcard = next((t for k, t in dict_entries if k == "*"), None)
-    return tuple(wildcard or ())
+    if stripped_matches:
+        return tuple(stripped_matches[-1])
+    wildcard_matches = tuple(t for k, t in dict_entries if k == "*")
+    if wildcard_matches:
+        return tuple(wildcard_matches[-1])
+    return ()
 
 
 def _candidate_model_groups(
diff --git a/tests/test_litellm/router_utils/test_fallback_event_handlers.py b/tests/test_litellm/router_utils/test_fallback_event_handlers.py
index f856c601d6b..ea8aa6e6a9a 100644
--- a/tests/test_litellm/router_utils/test_fallback_event_handlers.py
+++ b/tests/test_litellm/router_utils/test_fallback_event_handlers.py
@@ -275,6 +275,45 @@ def test_stripped_model_group_fallback_is_classified():
     assert PARTIAL in result[1]["content"]
 
 
+def test_duplicate_stripped_keys_pick_last_match():
+    """get_fallback_model_group keeps overwriting stripped matches without
+    breaking, so the last duplicate key in the list is the one the runtime
+    fallback hop calls. The capability scan must mirror that or it will miss
+    a prefill-rejecting target hidden behind a prefill-supporting first hit."""
+    result = build_mid_stream_continuation_messages(
+        messages=MESSAGES,
+        generated_content=PARTIAL,
+        model_group="openai/gpt-3.5-turbo",
+        fallbacks=[
+            {"gpt-3.5-turbo": ["gpt-4o"]},
+            {"gpt-3.5-turbo": ["claude-sonnet-4-6"]},
+        ],
+    )
+    assert len(result) == 2
+    assert result[1]["role"] == "user"
+    assert PARTIAL in result[1]["content"]
+    assert all(m.get("prefix") is not True for m in result)
+
+
+def test_duplicate_wildcard_keys_pick_last_match():
+    """Same last-wins semantics for wildcard entries: get_fallback_model_group
+    rewrites generic_fallback_idx for every ``*`` entry, so the last one is
+    the runtime target."""
+    result = build_mid_stream_continuation_messages(
+        messages=MESSAGES,
+        generated_content=PARTIAL,
+        model_group="some-unmatched-group",
+        fallbacks=[
+            {"*": ["gpt-4o"]},
+            {"*": ["claude-sonnet-4-6"]},
+        ],
+    )
+    assert len(result) == 2
+    assert result[1]["role"] == "user"
+    assert PARTIAL in result[1]["content"]
+    assert all(m.get("prefix") is not True for m in result)
+
+
 def test_exact_group_match_takes_priority_over_wildcard():
     """When both an exact group match and a wildcard fallback are configured,
     only the exact target is tried, so a prefill-rejecting model behind the