From 8dd25b9e661a7cc4ef76e72f693d33e1256889c4 Mon Sep 17 00:00:00 2001
From: LancyZhao <207653224+LancyZhao@users.noreply.github.com>
Date: Mon, 24 Aug 2026 17:16:04 +0800
Subject: [PATCH] fix(router): break cost-based-routing input+output ties on
 cache-read price

---
 litellm/router_strategy/lowest_cost.py        |  9 +++-
 .../local_testing/test_lowest_cost_routing.py | 49 +++++++++++++++++++
 2 files changed, 56 insertions(+), 2 deletions(-)

diff --git a/litellm/router_strategy/lowest_cost.py b/litellm/router_strategy/lowest_cost.py
index b927df0c438..cdcccf91921 100644
--- a/litellm/router_strategy/lowest_cost.py
+++ b/litellm/router_strategy/lowest_cost.py
@@ -262,6 +262,11 @@ class LowestCostLoggingHandler(CustomLogger):
             if item_output_cost is None:
                 item_output_cost = item_litellm_model_cost_map.get("output_cost_per_token", 5.0)
 
+            # Secondary ranking signal used to break ties on total input+output price.
+            # Fall back to the input cost when a model has no cache-read price, so models
+            # missing that field are not ranked as if their cache reads were free.
+            item_cache_read_cost = item_litellm_model_cost_map.get("cache_read_input_token_cost", item_input_cost)
+
             # if litellm["model"] is not in model_cost map -> use item_cost = $10
 
             item_cost = item_input_cost + item_output_cost
@@ -294,12 +299,12 @@ class LowestCostLoggingHandler(CustomLogger):
             ):  # if user passed in tpm / rpm in the model_list
                 continue
             else:
-                potential_deployments.append((_deployment, item_cost))
+                potential_deployments.append((_deployment, item_cost, item_cache_read_cost))
 
         if len(potential_deployments) == 0:
             return None
 
-        potential_deployments = sorted(potential_deployments, key=lambda x: x[1])
+        potential_deployments = sorted(potential_deployments, key=lambda x: (x[1], x[2]))
 
         selected_deployment: Final = potential_deployments[0][0]
         return selected_deployment
diff --git a/tests/local_testing/test_lowest_cost_routing.py b/tests/local_testing/test_lowest_cost_routing.py
index 5bf3a3ee98b..4ad33f8db42 100644
--- a/tests/local_testing/test_lowest_cost_routing.py
+++ b/tests/local_testing/test_lowest_cost_routing.py
@@ -199,3 +199,52 @@ async def test_get_available_endpoints_tpm_rpm_check_async(ans_rpm):
     assert (d_ans and d_ans["model_info"]["id"]) == ans
 
     print("selected deployment:", d_ans)
+
+
+@pytest.mark.parametrize("cheaper_cache_first", [True, False])
+@pytest.mark.asyncio
+async def test_lowest_cost_routing_breaks_input_output_tie_on_cache_read_cost(
+    cheaper_cache_first,
+):
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/38064
+
+    When two deployments tie on input + output price, cost-based routing must
+    fall back to the cache-read price instead of returning whichever deployment
+    happens to be listed first in the model_list.
+    """
+    import litellm
+
+    cheaper_cache = "tencent/deepseek-v4-flash"  # lower cache_read_input_token_cost
+    pricier_cache = "fireworks_ai/deepseek-v4-flash"
+
+    # The assertion is only meaningful while the shipped price map still ties these
+    # two models on input+output and separates them on cache-read price.
+    cheap = litellm.model_cost[cheaper_cache]
+    pricey = litellm.model_cost[pricier_cache]
+    assert (
+        cheap["input_cost_per_token"] + cheap["output_cost_per_token"]
+        == pricey["input_cost_per_token"] + pricey["output_cost_per_token"]
+    ), "precondition: the two deployments must tie on input+output price"
+    assert (
+        cheap["cache_read_input_token_cost"] < pricey["cache_read_input_token_cost"]
+    ), "precondition: cheaper_cache must have the lower cache-read price"
+
+    ordered = [cheaper_cache, pricier_cache] if cheaper_cache_first else [pricier_cache, cheaper_cache]
+    model_list = [
+        {
+            "model_name": "deepseek-v4-flash",
+            "litellm_params": {"model": model},
+            "model_info": {"id": model},
+        }
+        for model in ordered
+    ]
+
+    lowest_cost_logger = LowestCostLoggingHandler(router_cache=DualCache())
+
+    selected = await lowest_cost_logger.async_get_available_deployments(
+        model_group="deepseek-v4-flash", healthy_deployments=model_list
+    )
+
+    # Cheaper cache-read deployment must win regardless of model_list ordering.
+    assert selected["model_info"]["id"] == cheaper_cache