From a54319da431bf5e69c7bbb2c1eaf565113fb7fef Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Wed, 8 Jul 2026 17:06:44 +0000
Subject: [PATCH] fix(fireworks_ai): honor cached_tokens in cost calculator

The fireworks_ai cost calculator billed 100% of prompt tokens at the
full input rate, ignoring usage.prompt_tokens_details.cached_tokens and
the model's cache_read_input_token_cost. Split the prompt cost so cached
tokens are billed at the cache-read rate, falling back to the full input
rate when no cache-read price is set.

Fixes #32496
---
 litellm/llms/fireworks_ai/cost_calculator.py | 13 ++++-
 tests/local_testing/test_completion_cost.py  | 61 ++++++++++++++++++++
 2 files changed, 72 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/fireworks_ai/cost_calculator.py b/litellm/llms/fireworks_ai/cost_calculator.py
index ed936f6233a..ad5bb91bbc0 100644
--- a/litellm/llms/fireworks_ai/cost_calculator.py
+++ b/litellm/llms/fireworks_ai/cost_calculator.py
@@ -2,7 +2,7 @@
 For calculating cost of fireworks ai serverless inference models.
 """
 
-from typing import Tuple
+from typing import Optional, Tuple
 
 from litellm.constants import (
     FIREWORKS_AI_4_B,
@@ -75,8 +75,17 @@ def cost_per_token(model: str, usage: Usage) -> Tuple[float, float]:
         model_info = get_model_info(model=base_model, custom_llm_provider="fireworks_ai")
 
     ## CALCULATE INPUT COST
+    input_cost_per_token: float = model_info["input_cost_per_token"]
+    cache_read_cost: Optional[float] = model_info.get("cache_read_input_token_cost")
 
-    prompt_cost: float = usage["prompt_tokens"] * model_info["input_cost_per_token"]
+    cached_tokens: int = 0
+    if usage.prompt_tokens_details:
+        cached_tokens = usage.prompt_tokens_details.cached_tokens or 0
+
+    non_cached_tokens: int = max(usage["prompt_tokens"] - cached_tokens, 0)
+    prompt_cost: float = non_cached_tokens * input_cost_per_token + cached_tokens * (
+        cache_read_cost if cache_read_cost is not None else input_cost_per_token
+    )
 
     ## CALCULATE OUTPUT COST
     completion_cost = usage["completion_tokens"] * model_info["output_cost_per_token"]
diff --git a/tests/local_testing/test_completion_cost.py b/tests/local_testing/test_completion_cost.py
index cf0c645615d..0ac155d6de1 100644
--- a/tests/local_testing/test_completion_cost.py
+++ b/tests/local_testing/test_completion_cost.py
@@ -1225,6 +1225,67 @@ def test_completion_cost_fireworks_ai(model):
     assert cost > 0
 
 
+def test_fireworks_ai_cost_per_token_honors_cached_tokens():
+    """
+    Regression for https://github.com/BerriAI/litellm/issues/32496
+
+    The fireworks_ai cost calculator billed 100% of prompt tokens at the full
+    input rate, ignoring usage.prompt_tokens_details.cached_tokens and the
+    model's cache_read_input_token_cost. Cached tokens must be billed at the
+    cache-read rate.
+    """
+    from litellm.llms.fireworks_ai.cost_calculator import cost_per_token
+    from litellm.types.utils import PromptTokensDetailsWrapper, Usage
+
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    model = "accounts/fireworks/models/glm-5p2"
+    input_cost = litellm.model_cost["fireworks_ai/" + model]["input_cost_per_token"]
+    cache_read_cost = litellm.model_cost["fireworks_ai/" + model]["cache_read_input_token_cost"]
+    assert cache_read_cost < input_cost
+
+    usage = Usage(
+        prompt_tokens=8174,
+        completion_tokens=200,
+        total_tokens=8374,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=8165),
+    )
+
+    prompt_cost, _ = cost_per_token(model=model, usage=usage)
+
+    expected = (8174 - 8165) * input_cost + 8165 * cache_read_cost
+    assert prompt_cost == pytest.approx(expected)
+    # must be far below the old "everything at full input rate" behavior
+    assert prompt_cost < usage.prompt_tokens * input_cost
+
+
+def test_fireworks_ai_cost_per_token_no_cache_read_price_falls_back():
+    """
+    When a fireworks model has no cache_read_input_token_cost, cached tokens
+    fall back to the full input rate so behavior is unchanged for those models.
+    """
+    from litellm.llms.fireworks_ai.cost_calculator import cost_per_token
+    from litellm.types.utils import PromptTokensDetailsWrapper, Usage
+
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    model = "accounts/fireworks/models/deepseek-v3p1"
+    assert litellm.model_cost["fireworks_ai/" + model].get("cache_read_input_token_cost") is None
+    input_cost = litellm.model_cost["fireworks_ai/" + model]["input_cost_per_token"]
+
+    usage = Usage(
+        prompt_tokens=1000,
+        completion_tokens=50,
+        total_tokens=1050,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=800),
+    )
+
+    prompt_cost, _ = cost_per_token(model=model, usage=usage)
+    assert prompt_cost == pytest.approx(1000 * input_cost)
+
+
 def test_cost_azure_openai_prompt_caching():
     from litellm.utils import Choices, Message, ModelResponse, Usage
     from litellm.types.utils import (