Avoid double counting cache tokens in tier check

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -619,12 +619,20 @@ def _get_effective_prompt_tokens_for_tiered_pricing(usage: Usage) -> float:
     """
     if usage.prompt_tokens_details is not None:
         details = usage.prompt_tokens_details
-        text_tokens = float(getattr(details, "text_tokens", None) or 0)
-        if text_tokens == 0:
-            # text_tokens not set by this provider; fall back to prompt_tokens
-            text_tokens = float(getattr(usage, "prompt_tokens", 0) or 0)
+        raw_text_tokens = getattr(details, "text_tokens", None)
         cached_tokens = float(getattr(details, "cached_tokens", 0) or 0)
         cache_creation = float(getattr(details, "cache_creation_tokens", 0) or 0)
+        if raw_text_tokens is None:
+            # Some providers populate prompt_tokens_details without text_tokens.
+            # If prompt_tokens already covers the cache detail total, use it as
+            # the rolled-up input total. Otherwise cache tokens were reported
+            # outside prompt_tokens and must be added for the tier threshold.
+            prompt_tokens = float(getattr(usage, "prompt_tokens", 0) or 0)
+            cache_detail_tokens = cached_tokens + cache_creation
+            if cache_detail_tokens > 0 and prompt_tokens < cache_detail_tokens:
+                return prompt_tokens + cache_detail_tokens
+            return prompt_tokens
+        text_tokens = float(raw_text_tokens or 0)
         return text_tokens + cached_tokens + cache_creation
 
     # No prompt_tokens_details — add explicit cache fields only if they are

tests/test_litellm/test_cost_calculator.py

@@ -2594,6 +2594,73 @@ def test_custom_pricing_above_200k_cache_creation_tokens_uses_tiered_rate():
         litellm.model_cost.pop(f"vertex_ai/{model_id}", None)
 
 
+def test_custom_pricing_does_not_double_count_cached_tokens_when_text_tokens_missing():
+    """
+    Providers can send prompt_tokens_details without text_tokens while the rolled
+    up prompt_tokens value already includes cache tokens. Tier checks should not
+    add those cache tokens a second time.
+    """
+
+    model_id = "tiered-custom-pricing-anthropic-details-test"
+    model_info = {
+        "litellm_provider": "anthropic",
+        "mode": "chat",
+        "input_cost_per_token": 0.000005,
+        "output_cost_per_token": 0.000025,
+        "cache_creation_input_token_cost": 0.00000625,
+        "cache_read_input_token_cost": 0.0000005,
+        "input_cost_per_token_above_200k_tokens": 0.00001,
+        "output_cost_per_token_above_200k_tokens": 0.0000375,
+        "cache_creation_input_token_cost_above_200k_tokens": 0.0000125,
+        "cache_read_input_token_cost_above_200k_tokens": 0.000001,
+    }
+    cache_creation_tokens = 120_000
+
+    litellm.register_model(model_cost={model_id: model_info})
+    try:
+        response = ModelResponse(
+            id="test-id",
+            created=1234567890,
+            model="anthropic/claude-sonnet-4-5",
+            object="chat.completion",
+            choices=[],
+            usage=Usage(
+                prompt_tokens=130_000,
+                completion_tokens=0,
+                total_tokens=130_000,
+                prompt_tokens_details=PromptTokensDetailsWrapper(
+                    cache_creation_tokens=cache_creation_tokens
+                ),
+            ),
+        )
+
+        cost = completion_cost(
+            completion_response=response,
+            model="anthropic/claude-sonnet-4-5",
+            custom_llm_provider="anthropic",
+            custom_pricing=True,
+            router_model_id=model_id,
+        )
+
+        non_cache_prompt_tokens = 10_000
+        expected_cost = (
+            non_cache_prompt_tokens * model_info["input_cost_per_token"]
+            + cache_creation_tokens * model_info["cache_creation_input_token_cost"]
+        )
+        unexpected_tiered_cost = (
+            non_cache_prompt_tokens
+            * model_info["input_cost_per_token_above_200k_tokens"]
+            + cache_creation_tokens
+            * model_info["cache_creation_input_token_cost_above_200k_tokens"]
+        )
+
+        assert cost == pytest.approx(expected_cost)
+        assert cost != pytest.approx(unexpected_tiered_cost)
+    finally:
+        litellm.model_cost.pop(model_id, None)
+        litellm.model_cost.pop(f"anthropic/{model_id}", None)
+
+
 def test_extract_cache_read_tokens_anthropic_top_level():
     from litellm.proxy.spend_tracking.savings import extract_cache_read_tokens as _extract_cache_read_tokens