Merge pull request #41112 from BerriAI/litellm_registry_audit_2026_09_14

fix(models): rolling registry audit: Gemini latest aliases, Nova cache pricing, OpenRouter/Together sync, Mistral GLM 5.3, Azure snapshots, Grok caching

tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py

@@ -5,12 +5,15 @@ from typing import NamedTuple
 import pytest
 
 import litellm
+from litellm.cost_calculator import completion_cost
 from litellm.llms.bedrock.chat.converse_transformation import AmazonConverseConfig
 from litellm.llms.bedrock.common_utils import BedrockModelInfo
 from litellm.types.utils import (
     Choices,
     Message,
     ModelResponse,
+    PromptTokensDetailsWrapper,
+    Usage,
 )
 
 
@@ -152,3 +155,55 @@ def test_bedrock_gpt_5_6_offers_tools_and_reasoning_effort_but_not_thinking(prof
     assert "reasoning_effort" in supported
     assert "thinking" not in supported
     assert "output_config" not in supported
+
+
+# Cache-read prices are the `*-cache-read-input-tokens` usagetype rows of the AWS Price List API, us-east-1,
+# https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/us-east-1/index.json on 2026-09-15
+@pytest.mark.parametrize(
+    "model,expected_cache_read",
+    [
+        ("amazon.nova-lite-v1:0", 1.5e-8),
+        ("us.amazon.nova-lite-v1:0", 1.5e-8),
+        ("amazon.nova-micro-v1:0", 8.75e-9),
+        ("us.amazon.nova-micro-v1:0", 8.75e-9),
+        ("amazon.nova-pro-v1:0", 2e-7),
+        ("us.amazon.nova-pro-v1:0", 2e-7),
+        ("us.amazon.nova-premier-v1:0", 6.25e-7),
+    ],
+)
+def test_bedrock_nova_cache_read_prices(
+    model, expected_cache_read, local_model_cost_map
+):
+    model_info = litellm.model_cost[model]
+    assert model_info["cache_read_input_token_cost"] == expected_cache_read
+    usage = Usage(
+        prompt_tokens=1_000,
+        completion_tokens=100,
+        total_tokens=1_100,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=400),
+    )
+    response = _bedrock_response(model, usage)
+
+    cost = completion_cost(
+        completion_response=response,
+        model=model,
+        custom_llm_provider="bedrock",
+    )
+    expected_cost = (
+        600 * model_info["input_cost_per_token"]
+        + 400 * expected_cache_read
+        + 100 * model_info["output_cost_per_token"]
+    )
+    assert cost == pytest.approx(expected_cost)
+
+    uncached_usage = Usage(
+        prompt_tokens=1_000,
+        completion_tokens=100,
+        total_tokens=1_100,
+    )
+    uncached_cost = completion_cost(
+        completion_response=_bedrock_response(model, uncached_usage),
+        model=model,
+        custom_llm_provider="bedrock",
+    )
+    assert cost < uncached_cost

tests/test_litellm/llms/gemini/test_cost_calculator.py

@@ -1,8 +1,7 @@
-import os
-
 import pytest
 
 import litellm
+from litellm.cost_calculator import completion_cost
 from litellm.llms.gemini.cost_calculator import (
     cost_per_google_maps_grounding_request,
     cost_per_web_search_request,
@@ -18,6 +17,7 @@ from litellm.types.utils import (
     ImageResponse,
     ImageUsage,
     ImageUsageInputTokensDetails,
+    ModelResponse,
     PromptTokensDetailsWrapper,
     Usage,
 )
@@ -452,6 +452,42 @@ def test_map_traffic_type_to_service_tier(
     )
 
 
+# Alias targets are the `modelVersion` returned by
+# POST https://generativelanguage.googleapis.com/v1beta/models/<alias>:generateContent on 2026-09-15
+@pytest.mark.parametrize(
+    "alias,target",
+    [
+        ("gemini/gemini-flash-latest", "gemini/gemini-3.8-flash"),
+        ("gemini/gemini-flash-lite-latest", "gemini/gemini-3.5-flash-lite"),
+        ("gemini/gemini-pro-latest", "gemini/gemini-3.1-pro-preview"),
+    ],
+)
+def test_latest_aliases_cost_the_same_as_their_current_target(
+    monkeypatch, alias, target
+):
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+
+    usage = Usage(
+        prompt_tokens=1_000,
+        completion_tokens=500,
+        total_tokens=1_500,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=400),
+    )
+
+    def cost_of(model: str) -> float:
+        return completion_cost(
+            completion_response=ModelResponse(model=model, usage=usage),
+            model=model,
+            custom_llm_provider="gemini",
+        )
+
+    alias_cost = cost_of(alias)
+    target_cost = cost_of(target)
+    assert alias_cost == pytest.approx(target_cost)
+    assert alias_cost > 0
+
+
 @pytest.mark.parametrize(
     "prefixed,bare",
     [

tests/test_litellm/test_vertex_ai_xai_grok_prompt_caching_metadata.py

@@ -0,0 +1,38 @@
+from typing import Final
+
+import pytest
+
+import litellm
+from litellm import get_model_info
+from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
+from litellm.utils import supports_prompt_caching
+
+MODEL: Final = "vertex_ai/xai/grok-4.6"
+GROK_KEY_PREFIXES: Final = ("vertex_ai/xai/grok-", "azure_ai/grok-", "xai/grok-")
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+def test_grok_models_with_cache_read_price_advertise_prompt_caching() -> None:
+    cached_grok_models = tuple(
+        key
+        for key, entry in litellm.model_cost.items()
+        if key.startswith(GROK_KEY_PREFIXES) and entry.get("cache_read_input_token_cost")
+    )
+    assert cached_grok_models, "expected at least one grok model with a cache read price"
+
+    missing_flag = tuple(key for key in cached_grok_models if supports_prompt_caching(model=key) is not True)
+    assert missing_flag == (), (
+        f"grok models with cache_read_input_token_cost fail supports_prompt_caching: {missing_flag}"
+    )
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+def test_vertex_ai_grok_4_6_supports_prompt_caching_via_get_model_info() -> None:
+    routed_model, provider, _, _ = get_llm_provider(model=MODEL)
+    assert (routed_model, provider) == ("xai/grok-4.6", "vertex_ai")
+
+    info = get_model_info(model=routed_model, custom_llm_provider=provider)
+    assert info["litellm_provider"] == "vertex_ai"
+    assert info.get("supports_prompt_caching") is True
+
+    assert supports_prompt_caching(model=MODEL) is True