fix(cost_calculator): resolve real cost key when model_name alias contains '/'

When a router-facing model_name alias contains a '/' whose leading segment
is not a registered provider (e.g. 'vertex/claude-opus-5' for deployment
'vertex_ai/claude-opus-5'), _select_model_name_for_cost_calc re-prefixed
it into a non-existent key ('vertex_ai/vertex/claude-opus-5'), so cost
lookup silently priced every streamed request at $0 - token counts were
recorded, no error raised, budgets never tripped.

After prefixing, walk the alias tail and return the first assembly that
exists in litellm.model_cost ('vertex_ai/claude-opus-5'). Provider/region
segments in the head are preserved, and an alias that resolves to no
known key keeps the previous behavior (no crash, legacy double-prefix).

Fixes #38069

litellm/cost_calculator.py

@@ -799,9 +799,52 @@ def _select_model_name_for_cost_calc(
         else:
             return_model = f"{custom_llm_provider}/{return_model}"
 
+        # A router-facing model_name alias may itself contain a "/" (e.g. "vertex/claude-opus-5")
+        # whose leading segment is NOT a registered provider, so the alias was re-prefixed above
+        # into a non-existent key like "vertex_ai/vertex/claude-opus-5". Strip unregistered
+        # leading segments until the assembled name is a known model_cost key (or nothing
+        # strippable remains), so cost lookup resolves to the real model instead of pricing
+        # the request at $0. See #38069.
+        stripped = _strip_unregistered_leading_segments(return_model, region_name)
+        if stripped is not None:
+            return_model = stripped
+
     return return_model
 
 
+def _strip_unregistered_leading_segments(model: str, region_name: str | None) -> str | None:
+    """Find the real cost-map key hidden inside a re-prefixed alias.
+
+    After the provider (and optional region) was prepended to a router-facing alias that
+    itself contains "/" segments (e.g. "vertex_ai/vertex/claude-opus-5"), walk the
+    remaining segments: keep the provider/region head, then try successively shorter
+    tails — "vertex_ai/vertex/claude-opus-5", "vertex_ai/claude-opus-5" — returning the
+    first that exists in `litellm.model_cost`. Segments that are themselves providers or
+    regions are never dropped from the head, so "vertex_ai/us-east-1/model" keeps its region.
+
+    Returns None if no candidate resolves; the caller keeps its original assembly.
+    """
+
+    segments = model.split("/")
+    # head = leading provider (+ optional region) that must be preserved
+    head_len = 1
+    if region_name is not None and len(segments) > 1 and segments[1] == region_name:
+        head_len = 2
+
+    head = segments[:head_len]
+    tail = segments[head_len:]
+    while tail:
+        candidate = "/".join(head + tail)
+        if candidate in litellm.model_cost:
+            return candidate
+        # only strip a leading tail segment that is NOT itself meaningful:
+        # a provider/region segment inside the tail (e.g. "azure_ai/o3") stays
+        if tail[0] in LlmProvidersSet or tail[0] == region_name:
+            return None
+        tail = tail[1:]
+    return None
+
+
 @lru_cache(maxsize=DEFAULT_MAX_LRU_CACHE_SIZE)
 def _model_contains_known_llm_provider(model: str) -> bool:
     """

tests/test_litellm/test_cost_calculator.py

@@ -3781,3 +3781,88 @@ def test_completion_cost_prices_anthropic_shaped_cache_read_tokens(_local_model_
     )
 
     assert cost == pytest.approx(3 * 4e-6 + 4014 * 4e-7 + 5 * 2e-5, rel=1e-9)
+
+
+def test_select_model_name_strips_unregistered_alias_prefix(_local_model_cost_map):
+    """A router-facing model_name alias containing "/" whose leading segment is NOT a
+    registered provider must not be double-prefixed into a non-existent cost key.
+
+    Regression test for #38069: alias "vertex/claude-opus-5" (real deployment
+    "vertex_ai/claude-opus-5") was re-prefixed into "vertex_ai/vertex/claude-opus-5",
+    silently pricing every streamed request at $0.
+    """
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="vertex/claude-opus-5",
+    )
+    response._hidden_params = {}
+
+    selected = _select_model_name_for_cost_calc(
+        model=None,
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert selected == "vertex_ai/claude-opus-5"
+
+
+def test_completion_cost_nonzero_for_slash_alias_model_name(_local_model_cost_map):
+    """End-to-end cost through a "/"-containing alias must price above zero (#38069)."""
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="vertex/claude-opus-5",
+    )
+    response._hidden_params = {"custom_llm_provider": "vertex_ai"}
+    response.usage = litellm.Usage(prompt_tokens=100, completion_tokens=50)
+
+    cost = litellm.completion_cost(
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert cost == pytest.approx(100 * 5e-6 + 50 * 2.5e-5, rel=1e-9)
+
+
+def test_select_model_name_unresolvable_alias_unchanged(_local_model_cost_map):
+    """An alias that resolves to no known cost key keeps the legacy double-prefixed name."""
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="team/nonsense-model",
+    )
+    response._hidden_params = {}
+
+    selected = _select_model_name_for_cost_calc(
+        model=None,
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert selected == "vertex_ai/team/nonsense-model"