fix(cost): preserve Claude generation suffixes in Vertex model strip

_strip_stable_vertex_version used -\\d+$ which rewrote claude-sonnet-5
to claude-sonnet while leaving claude-sonnet-5@default unchanged, so bare
and @default Vertex Anthropic ids diverged in cost-map lookup (issue
#35758). Only strip 3+ digit publish ids / 8-digit dates. Also form
vertex_ai/ cost-map keys when the provider tag is vertex_ai-*.

litellm/utils.py

@@ -4894,8 +4894,31 @@ def get_max_tokens(model: str) -> int | None:
         )
 
 
-def _strip_stable_vertex_version(model_name) -> str:
-    return re.sub(r"-\d+$", "", model_name)
+def _strip_stable_vertex_version(model_name: str) -> str:
+    """Strip Vertex AI *publish* version suffixes, not product generation numbers.
+
+    Vertex stable/versioned ids look like ``gemini-1.5-flash-001`` or
+    ``gemini-2.0-flash-001`` (3+ digit build ids). Anthropic-on-Vertex product
+    names use short generation numbers (``claude-sonnet-5``, ``claude-opus-4-1``)
+    and optional region resource names (``claude-sonnet-5@default``).
+
+    The old ``-\\d+$`` pattern treated ``-5`` as a version and rewrote
+    ``claude-sonnet-5`` → ``claude-sonnet``, while leaving
+    ``claude-sonnet-5@default`` untouched. That made bare and ``@default``
+    cost-map lookups diverge (issue #35758): bare ids missed priced entries
+    and logged $0, while ``@default`` still resolved.
+
+    Only strip:
+    - 3+ digit trailing version codes (``-001``, ``-002``)
+    - 8-digit date suffixes (``-20250929``)
+    Leave region suffixes after ``@`` alone.
+    """
+    if not model_name:
+        return model_name
+    if "@" in model_name:
+        # Region / resource name (e.g. model@default) is not a publish version.
+        return model_name
+    return re.sub(r"-(?:\d{8}|\d{3,})$", "", model_name)
 
 
 def _get_base_bedrock_model(model_name) -> str:
@@ -5159,6 +5182,23 @@ def _get_model_info_from_generalization(
     return None
 
 
+def _cost_map_provider_prefix(custom_llm_provider: str | None) -> str | None:
+    """Map sub-provider tags to the cost-map key prefix.
+
+    Cost-map keys use ``vertex_ai/...`` even when ``litellm_provider`` is the
+    more specific ``vertex_ai-anthropic_models`` (or other ``vertex_ai-*``
+    tags). Using the sub-provider as a key prefix produces miss keys like
+    ``vertex_ai-anthropic_models/claude-sonnet-5`` and falls through to $0.
+    """
+    if custom_llm_provider is None:
+        return None
+    if custom_llm_provider.startswith("vertex_ai"):
+        return "vertex_ai"
+    if custom_llm_provider.startswith("fireworks_ai"):
+        return "fireworks_ai"
+    return custom_llm_provider
+
+
 def _get_potential_model_names(model: str, custom_llm_provider: str | None) -> PotentialModelNamesAndCustomLLMProvider:
     if custom_llm_provider is None:
         # Get custom_llm_provider
@@ -5170,18 +5210,21 @@ def _get_potential_model_names(model: str, custom_llm_provider: str | None) -> P
         combined_model_name = model
         stripped_model_name = _strip_model_name(model=model, custom_llm_provider=custom_llm_provider)
         combined_stripped_model_name = stripped_model_name
-    elif custom_llm_provider and model.startswith(
-        custom_llm_provider + "/"
-    ):  # handle case where custom_llm_provider is provided and model starts with custom_llm_provider
-        split_model = model.split("/", 1)[1]
-        combined_model_name = model
-        stripped_model_name = _strip_model_name(model=split_model, custom_llm_provider=custom_llm_provider)
-        combined_stripped_model_name = f"{custom_llm_provider}/{stripped_model_name}"
     else:
-        split_model = model
-        combined_model_name = f"{custom_llm_provider}/{model}"
-        stripped_model_name = _strip_model_name(model=model, custom_llm_provider=custom_llm_provider)
-        combined_stripped_model_name = f"{custom_llm_provider}/{stripped_model_name}"
+        cost_map_prefix: Final = _cost_map_provider_prefix(custom_llm_provider) or custom_llm_provider
+        # Prefer the cost-map prefix (e.g. vertex_ai/) over sub-provider tags when
+        # stripping a leading provider segment from the model string.
+        if model.startswith(custom_llm_provider + "/"):
+            split_model = model.split("/", 1)[1]
+            combined_model_name = f"{cost_map_prefix}/{split_model}"
+        elif cost_map_prefix != custom_llm_provider and model.startswith(cost_map_prefix + "/"):
+            split_model = model.split("/", 1)[1]
+            combined_model_name = model if model.startswith(cost_map_prefix + "/") else f"{cost_map_prefix}/{split_model}"
+        else:
+            split_model = model
+            combined_model_name = f"{cost_map_prefix}/{model}"
+        stripped_model_name = _strip_model_name(model=split_model, custom_llm_provider=cost_map_prefix)
+        combined_stripped_model_name = f"{cost_map_prefix}/{stripped_model_name}"
 
     if custom_llm_provider in ("bedrock", "bedrock_converse"):
         from litellm.llms.bedrock.common_utils import strip_bedrock_routing_prefix

tests/test_litellm/test_vertex_claude_sonnet5_cost_parity.py

@@ -0,0 +1,149 @@
+"""
+Regression for https://github.com/BerriAI/litellm/issues/35758
+
+``vertex_ai/claude-sonnet-5`` must cost the same as ``vertex_ai/claude-sonnet-5@default``.
+
+Root cause: ``_strip_stable_vertex_version`` treated the product generation
+suffix ``-5`` as a Vertex publish version and rewrote bare ids to
+``claude-sonnet``, while ``@default`` resource names were left intact.
+"""
+
+from __future__ import annotations
+
+import os
+
+import pytest
+
+import litellm
+from litellm.cost_calculator import completion_cost, cost_per_token
+from litellm.types.utils import Choices, Message, ModelResponse, Usage
+from litellm.utils import (
+    _get_potential_model_names,
+    _strip_model_name,
+    _strip_stable_vertex_version,
+    get_model_info,
+)
+
+
+@pytest.fixture
+def local_model_cost_map(monkeypatch):
+    """Use the bundled cost map so tests do not depend on network-fetched main."""
+    original = litellm.model_cost
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+    litellm.get_model_info.cache_clear()
+    try:
+        yield
+    finally:
+        litellm.model_cost = original
+        litellm.get_model_info.cache_clear()
+
+
+@pytest.mark.parametrize(
+    "model_name,expected",
+    [
+        # Product generation numbers must be preserved
+        ("claude-sonnet-5", "claude-sonnet-5"),
+        ("claude-sonnet-5@default", "claude-sonnet-5@default"),
+        ("claude-sonnet-4-5", "claude-sonnet-4-5"),
+        ("claude-opus-4-1", "claude-opus-4-1"),
+        # Real Vertex publish / date version suffixes still strip
+        ("gemini-1.5-flash-001", "gemini-1.5-flash"),
+        ("gemini-2.0-flash-001", "gemini-2.0-flash"),
+        ("some-model-20250929", "some-model"),
+    ],
+)
+def test_strip_stable_vertex_version_preserves_generation_numbers(model_name, expected):
+    assert _strip_stable_vertex_version(model_name) == expected
+    assert _strip_model_name(model_name, "vertex_ai") == expected
+
+
+def test_potential_names_for_vertex_anthropic_subprovider(local_model_cost_map):
+    """``vertex_ai-anthropic_models`` must still form ``vertex_ai/...`` cost keys."""
+    names = _get_potential_model_names(
+        model="claude-sonnet-5",
+        custom_llm_provider="vertex_ai-anthropic_models",
+    )
+    assert names["combined_model_name"] == "vertex_ai/claude-sonnet-5"
+    assert names["stripped_model_name"] == "claude-sonnet-5"
+    assert names["combined_stripped_model_name"] == "vertex_ai/claude-sonnet-5"
+
+
+def test_get_model_info_parity_bare_vs_default(local_model_cost_map):
+    bare = get_model_info(model="vertex_ai/claude-sonnet-5")
+    default = get_model_info(model="vertex_ai/claude-sonnet-5@default")
+    assert bare["input_cost_per_token"] == default["input_cost_per_token"]
+    assert bare["output_cost_per_token"] == default["output_cost_per_token"]
+    assert bare["input_cost_per_token"] and bare["input_cost_per_token"] > 0
+    # Provider-only bare id must resolve via vertex_ai cost-map prefix
+    via_sub = get_model_info(
+        model="claude-sonnet-5",
+        custom_llm_provider="vertex_ai-anthropic_models",
+    )
+    assert via_sub["input_cost_per_token"] == bare["input_cost_per_token"]
+    assert via_sub["key"] == "vertex_ai/claude-sonnet-5"
+
+
+def _usage() -> Usage:
+    return Usage(
+        prompt_tokens=1000,
+        completion_tokens=500,
+        total_tokens=1500,
+        cache_creation_input_tokens=100,
+        cache_read_input_tokens=200,
+    )
+
+
+@pytest.mark.parametrize(
+    "model",
+    [
+        "vertex_ai/claude-sonnet-5",
+        "vertex_ai/claude-sonnet-5@default",
+    ],
+)
+def test_completion_cost_nonzero_for_vertex_sonnet5(model, local_model_cost_map):
+    resp = ModelResponse(
+        model=model,
+        choices=[Choices(message=Message(role="assistant", content="ok"), finish_reason="stop")],
+        usage=_usage(),
+    )
+    cost = completion_cost(
+        completion_response=resp,
+        model=model,
+        custom_llm_provider="vertex_ai",
+    )
+    assert cost > 0
+
+
+def test_completion_cost_parity_bare_vs_default(local_model_cost_map):
+    usage = _usage()
+    costs = []
+    for model in ("vertex_ai/claude-sonnet-5", "vertex_ai/claude-sonnet-5@default"):
+        resp = ModelResponse(
+            model=model.split("/", 1)[1],
+            choices=[Choices(message=Message(role="assistant", content="ok"), finish_reason="stop")],
+            usage=usage,
+        )
+        costs.append(
+            completion_cost(
+                completion_response=resp,
+                model=model,
+                custom_llm_provider="vertex_ai",
+            )
+        )
+    assert costs[0] == costs[1]
+    assert costs[0] > 0
+
+
+def test_cost_per_token_with_vertex_anthropic_provider_tag(local_model_cost_map):
+    """Provider tag from the cost map must not zero out bare Claude-on-Vertex ids."""
+    prompt, completion = cost_per_token(
+        model="claude-sonnet-5",
+        prompt_tokens=1000,
+        completion_tokens=500,
+        cache_creation_input_tokens=100,
+        cache_read_input_tokens=200,
+        custom_llm_provider="vertex_ai-anthropic_models",
+        usage_object=_usage(),
+    )
+    assert prompt + completion > 0