From e696ab1850c6703c4e77e91b4ac7d48a0ac9a5f8 Mon Sep 17 00:00:00 2001
From: LancyZhao <207653224+LancyZhao@users.noreply.github.com>
Date: Wed, 2 Sep 2026 15:58:34 +0800
Subject: [PATCH] test(router): lock deployment-level cache-read price of 0 as
 a real price

a168bab rewrote the deployment-level lookup to drop two mutable-dict literals
and, in doing so, replaced a truthiness check with an explicit `is not None`.
That silently fixed a behavior bug the commit message never claimed: a
deployment configuring `cache_read_input_token_cost: 0` was read as unset, fell
through to the cost map and then to the input-cost fallback, and lost the tie to
a deployment whose cache reads are actually billed. Providers that do not charge
for cache reads make 0 a real configuration, not an edge case.

No test covered it, so nothing stops the next rewrite from going back to `or` /
truthiness. This adds the missing case: verified failing on bdfb64c's
implementation (both list orders) and passing on the current one.
---
 .../router_strategy/test_lowest_cost.py       | 41 +++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/tests/test_litellm/router_strategy/test_lowest_cost.py b/tests/test_litellm/router_strategy/test_lowest_cost.py
index 309780f217d..c95b16daeba 100644
--- a/tests/test_litellm/router_strategy/test_lowest_cost.py
+++ b/tests/test_litellm/router_strategy/test_lowest_cost.py
@@ -45,3 +45,44 @@ async def test_cost_routing_breaks_input_output_tie_on_cache_read_cost(cheaper_c
     )
 
     assert selected["model_info"]["id"] == "cheaper-cache"
+
+
+@pytest.mark.parametrize("free_cache_first", [True, False])
+@pytest.mark.asyncio
+async def test_cost_routing_honors_zero_deployment_cache_read_cost(free_cache_first):
+    """
+    A deployment-level cache_read_input_token_cost of 0 is a price, not a missing value.
+
+    Truthiness checks treat it as unset and fall through to the cost map / input-cost
+    fallback, which ranks a deployment whose cache reads are free as the priciest one
+    in a tie. Providers that do not charge for cache reads make this a real config.
+    """
+    free = {
+        "model_name": "cache-zero-test",
+        "litellm_params": {
+            "model": "openai/zero-model-not-in-cost-map",
+            "input_cost_per_token": 1e-06,
+            "output_cost_per_token": 2e-06,
+            "cache_read_input_token_cost": 0.0,
+        },
+        "model_info": {"id": "free-cache"},
+    }
+    paid = {
+        "model_name": "cache-zero-test",
+        "litellm_params": {
+            "model": "openai/zero-model-not-in-cost-map",
+            "input_cost_per_token": 1e-06,
+            "output_cost_per_token": 2e-06,
+            "cache_read_input_token_cost": 1e-08,
+        },
+        "model_info": {"id": "paid-cache"},
+    }
+    model_list = [free, paid] if free_cache_first else [paid, free]
+
+    logger = LowestCostLoggingHandler(router_cache=DualCache())
+
+    selected = await logger.async_get_available_deployments(
+        model_group="cache-zero-test", healthy_deployments=model_list
+    )
+
+    assert selected["model_info"]["id"] == "free-cache"